Last 7 Days (July 16 – July 22, 2026)
Skills are a useful abstraction for software agents, turning human and agent experience into reusable procedural knowledge. Yet existing skill libraries are mostly hand-written, text-centric, or derived from agent traces, leaving tutorial videos and other multimodal human resources largely underused. We present RESOURCE2SKILL, a framework that distills multimodal resources, including tutorial videos, repositories, articles, and reference artifacts, into executable skills for software agents. RESOURCE2SKILL organizes these skills as a hierarchical multimodal Skill Wiki, where each entry combines structured text, code, visual examples, metadata, and provenance. This design preserves complementary signals from different resources: videos capture temporal operations and visual effects, code captures executable tool patterns, and articles or artifacts provide conceptual and stylistic grounding. At inference time, agents retrieve and compose relevant skills from the wiki; when coverage is insufficient, the same construction operator can acquire new skills online. Across seven practical authoring domains, RESOURCE2SKILL improves average overall score by +11.9 percentage points over no-skill agents and outperforms strong harness baselines in 26 of 28 main-aggregate model-domain cells. Ablations confirm the value of multimodal skill format, hierarchical organization, source diversity, selection strategy, and online acquisition.
Primary: Microsoft Research
All Institutions: Microsoft Research, University of California, Santa Cruz, Shanghai Jiao Tong University
RESOURCE2SKILL presents a rigorous and impactful framework for distilling multimodal human resources into executable agent skills, demonstrating substantial performance gains across diverse software domains through a structured, hierarchical skill wiki and controlled online acquisition.
The paper introduces RESOURCE2SKILL, a framework for distilling multimodal human resources (videos, code repos, articles, artifacts) into a structured, hierarchical "Skill Wiki" for software agents. The core innovation lies in the schema: skills are tuples of text, visual, code, and metadata, organized by a domain-specific taxonomy. The method employs a "construction operator" that uses a vision-capable LLM to extract and normalize these components, followed by deterministic acceptance predicates (completeness, provenance, deduplication, modality consistency, structural executability). At inference, a two-stage retrieval process (lexical BM25 over taxonomy paths, then LLM selection) allows agents to compose skills. A key feature is the ability to trigger online acquisition when offline skills are insufficient, using the same construction operator. The approach is well-motivated, addressing the gap between implicit agent knowledge and explicit, reusable procedural knowledge derived from rich human resources.
The evaluation is extensive, covering seven diverse authoring domains (PPT, CAD, Web, Excel, Blender, UE5, Reaper) and four agent backends (GPT-5.5, GPT-5.4, Mini, Nano). The primary metric is an automated judge score (vision/audio models) on rendered artifacts, validated by human A/B studies. RESOURCE2SKILL shows significant improvements (+11.9 pp) over no-skill baselines and outperforms strong harness baselines (ClaudeCode-H, Codex-H) in 26/28 cells. Ablations confirm the value of multimodal content, hierarchical organization, and online acquisition. The scaling study shows performance saturates around 200 skills. The results are robust and statistically significant.
The paper provides a code link (https://aka.ms/Resource2Skill) and detailed appendices describing the benchmark construction, judge reliability, and experimental setup. The acceptance predicates are deterministic, and the library construction is frozen before benchmarking, ensuring no data leakage. The matched-brief design and fixed seeds enhance reproducibility. The detailed description of the MCP interface and domain backends allows for replication.
The framework relies on the availability of high-quality multimodal resources (videos, repos) for each domain, which may not exist for all software domains. The automated judge, while correlated with humans, may have biases (e.g., rewarding visual polish over functional correctness in some cases). The online acquisition adds latency. The "structural executability" check ensures code runs but does not guarantee it solves the specific task correctly, leading to potential partial grounding failures as noted in case studies.
This work significantly advances the field of software agents by providing a scalable method for building and maintaining skill libraries from diverse human-generated content. It enables agents to leverage existing human knowledge (tutorials, code) more effectively, potentially lowering the barrier to entry for complex software tasks. The hierarchical wiki structure offers a promising paradigm for organizing agent memory. RESOURCE2SKILL presents a rigorous and impactful framework for distilling multimodal human resources into executable agent skills, demonstrating substantial performance gains across diverse software domains through a structured, hierarchical skill wiki and controlled online acquisition.
Video models are evolving into vision foundation models, yet they still lack human-like multi-step reasoning. Streaming autoregressive diffusion models are efficient but limited in reasoning, while bidirectional diffusion enables global revision with high inference costs due to dense frame-level denoising. Both paradigms struggle to achieve logical consistency and low-latency streaming for complex reasoning tasks. We propose HDR (Hierarchical Denoising for Visual Reasoning), a unified framework that integrates hierarchical latents into causal video generation for multi-step reasoning. HDR organizes video latents into a tree-structured hierarchy, enabling coarse-to-fine reasoning before streaming output. Coarse denoising layers preserve uncertain hypotheses for global planning, while finer layers progressively refine them into concrete visual states. A sparse hierarchical attention pattern (SHAP) further reduces temporal attention costs. We introduce a level-stratified multi-step video reasoning benchmark with out-of-distribution cases, covering six tasks: maze navigation, Tower of Hanoi, one-line drawing, sliding puzzle, Sokoban, and water pouring. Compared with streaming autoregressive diffusion baselines, HDR improves success from 34.22 to 60.29 (76.2% relative gain) and increases average progress from 76.00 to 89.56, demonstrating more consistent reasoning trajectories. HDR maintains low-latency streaming at 0.70 seconds per latent, achieving 54.2 times faster inference than bidirectional diffusion. It also retains 82.9% of full-data performance with only 2% training data, compared with 52.0% for bidirectional diffusion. Real-world robot experiments further demonstrate HDR's potential for physical interaction and world modeling. Project demo: https://hierarchical-diffusion-reasoning.github.io/.
Primary: The Hong Kong University of Science and Technology
All Institutions: State Key Laboratory of Multimedia Information Processing, The Hong Kong University of Science and Technology, Beihang University, Fuzhou University, Peking University, School of Computer Science
HDR makes a significant contribution to the field of generative AI, particularly in video modeling and embodied AI. 1. **Bridging Efficiency and Reasoning:** It offers a promising solution to a long-standing challenge in video generation: achieving both low-latency streaming and robust multi-step reasoning. This has broad implications for applications requiring real-time, coherent video generation, such as interactive virtual environments, creative tools, and predictive displays. 2. **Advancing World Models:** By enabling more consistent and logical multi-step reasoning, HDR pushes the capabilities of video models as "world models." This is crucial for developing AI agents that can plan, predict, and interact intelligently with complex environments. 3. **Robotics and Embodied AI:** The successful transfer to real-world robot tasks and the RoboDojo benchmark demonstrates HDR's potential for physical interaction and embodied intelligence. Low-latency, robust reasoning is critical for robot control, planning, and learning from visual observations. 4. **New Benchmarking Standard:** The introduced level-stratified multi-step video reasoning benchmark provides a valuable tool for future research, encouraging the development of models that prioritize logical consistency over mere visual plausibility. 5. **Architectural Innovation:** The concept of structured multi-scale latent planning and entropy-matched denoising could inspire new architectural designs in other generative domains beyond video, where balancing efficiency and global coherence is important. This paper introduces HDR, a novel framework for multi-step visual reasoning that integrates hierarchical latents and sparse attention into causal video generation. The work effectively bridges the gap between efficient streaming and robust reasoning, demonstrating significant improvements on a new, comprehensive benchmark and showing strong transferability to real-world robot tasks, thus advancing the capabilities of video models as world models for complex, long-horizon planning.
The paper proposes Hierarchical Denoising for Visual Reasoning (HDR), a novel framework that addresses the fundamental tension between efficient streaming generation and robust multi-step reasoning in video models. The core idea is to integrate hierarchical latents into a causal video generation process, enabling coarse-to-fine reasoning before streaming output. This is a well-motivated approach, as existing autoregressive models struggle with revision, while bidirectional models are computationally expensive. The methodology introduces several key components: 1. **Tree-structured Hierarchy:** Video latents are organized into a multi-level hierarchy, where coarse tokens summarize global temporal structure (high-level plans) and fine tokens encode local visual details. This provides an explicit intermediate space for planning. 2. **Layer-wise Flow-Matching Objective:** The model is trained with a flow-matching objective across all hierarchy levels. A crucial design choice is the "hierarchy-matched denoising schedule," which assigns level-dependent sampling budgets, allowing coarse layers to maintain higher noise levels (preserving uncertainty for multiple global plans) and finer layers to progressively refine these hypotheses. This entropy-matched approach is a clever detail that aligns denoising strength with the abstraction level. 3. **Sparse Hierarchical Attention Pattern (SHAP):** To maintain efficiency, SHAP defines a structured attention mask over flattened tree tokens. Each token attends only to fixed local, parent-level, and first-frame contexts. This enables multi-scale information flow with reduced temporal attention cost, preserving streaming generation properties. 4. **HDR-WAM for Robotics:** The paper extends HDR to a World Action Model (HDR-WAM) for embodied control, combining episode-level visual context with local action-conditioned rollouts. This demonstrates the generality of the hierarchical approach. The methodology is sound, well-articulated, and directly tackles the identified problem. The combination of hierarchical latents, a tailored denoising schedule, and sparse attention is a strong technical contribution.
The experimental evaluation is comprehensive and rigorous, supporting the claims effectively. 1. **Novel Benchmark:** A significant contribution is the construction of a level-stratified multi-step video reasoning benchmark covering six diverse tasks (maze navigation, Tower of Hanoi, one-line drawing, sliding puzzle, Sokoban, water pouring). This benchmark includes out-of-distribution (OOD) cases and uses robust metrics (success and average progress), which are crucial for evaluating logical consistency over long trajectories. This benchmark itself is a valuable asset for the community. 2. **Strong Baselines:** HDR is compared against relevant full-attention baselines (Bidirectional Diffusion, VideoMAE) and streaming autoregressive baselines (CausalForcing, VideoGPT). The main diffusion baselines are built on the same foundation model (Wan2.2-5B-TI2V) for a controlled comparison. 3. **Quantitative Results:** HDR significantly outperforms the streaming autoregressive baseline (CausalForcing), improving overall success from 34.22 to 60.29 (76.2% relative gain) and average progress from 76.00 to 89.56. It achieves this while maintaining low-latency streaming (0.70s per latent, comparable to CausalForcing's 0.72s) and being substantially faster than bidirectional diffusion (37.92s). It also remains competitive with bidirectional diffusion in reasoning performance, especially considering the latency difference. 4. **Qualitative Evidence:** Visual comparisons demonstrate HDR's ability to perform hierarchical planning and avoid early local commitments that lead to failure in baselines. 5. **Ablation Studies:** * **Layer Importance:** Demonstrates that increasing the number of active hierarchy layers progressively improves performance, validating the coarse-to-fine reasoning process. * **Denoising-step Reduction:** HDR shows superior robustness under reduced denoising budgets compared to both streaming AR and bidirectional diffusion, highlighting the stability provided by coarse layers. * **Data Reduction:** HDR degrades more gracefully than bidirectional diffusion with limited training data, suggesting better data efficiency for learning reasoning rules. * **Entropy-Matched vs. Fully Denoised:** The ablation on denoising schedules confirms the benefit of the entropy-matched approach, which preserves uncertainty at coarse levels, over a naive full denoising strategy. 6. **Real-world Robot Experiments:** The transferability of HDR is demonstrated in a physical robot maze task and on the RoboDojo simulation benchmark. HDR-WAM shows strong performance, particularly on Long-Horizon and Memory tasks, without robot-domain pretraining, indicating its potential for physical interaction and world modeling. This is a very strong point, showcasing practical relevance. The experimental section is exceptionally thorough, covering performance, efficiency, robustness, and real-world applicability.
The paper provides a good level of detail for reproducibility. * The methodology is clearly described, including the hierarchical latent representation, layer-wise flow-matching objective, and SHAP. * The new benchmark is detailed, including task descriptions, difficulty stratification, OOD cases, and evaluation metrics. Appendix provides full benchmark construction and task-specific evaluation details. * Baselines, training details, and implementation settings are specified, including the base diffusion model (Wan2.2-5B-TI2V) and the number of hierarchy levels and denoising schedule. * The entropy-matched denoising schedule derivation is provided in the appendix. * Details for HDR-WAM (temporal views, action alignment, token layout, attention mask, RoboDojo data processing) are also in the appendix. The project demo page (https://hierarchical-diffusion-reasoning.github.io/) further suggests that code or interactive demonstrations might be available, enhancing reproducibility. Overall, the paper provides sufficient information for a skilled researcher to reproduce the main results.
1. **Fixed Hierarchy Structure:** The paper uses a fixed, tree-structured hierarchy. While effective, the optimal hierarchy structure might be task-dependent or could potentially be learned. The current approach might not be optimal for all types of reasoning problems or video lengths. 2. **Pre-computation Phase:** While HDR enables "streaming output," it still involves a coarse-to-fine reasoning phase *before* the final streaming output. This means there's an initial latency for the hierarchical planning, even if subsequent frame generation is fast. For truly real-time, continuous streaming where no look-ahead is possible, this might still be a limitation compared to purely causal models. 3. **Scalability to Extremely Long Videos:** While SHAP reduces attention cost, the total number of hierarchical tokens still grows with video length. For extremely long, open-ended video generation, the fixed-size local and parent contexts might eventually become insufficient for maintaining global consistency over vast temporal spans. 4. **Training Complexity:** Training a multi-level hierarchical diffusion model with a layer-wise objective and specific denoising schedules might be more complex and resource-intensive than training a single-level autoregressive model. 5. **Generalizability of Reasoning:** While the benchmark is diverse, it's still a set of structured reasoning tasks. The extent to which HDR's reasoning capabilities generalize to more open-ended, less structured, or abstract reasoning scenarios remains to be fully explored.
HDR makes a significant contribution to the field of generative AI, particularly in video modeling and embodied AI. 1. **Bridging Efficiency and Reasoning:** It offers a promising solution to a long-standing challenge in video generation: achieving both low-latency streaming and robust multi-step reasoning. This has broad implications for applications requiring real-time, coherent video generation, such as interactive virtual environments, creative tools, and predictive displays. 2. **Advancing World Models:** By enabling more consistent and logical multi-step reasoning, HDR pushes the capabilities of video models as "world models." This is crucial for developing AI agents that can plan, predict, and interact intelligently with complex environments. 3. **Robotics and Embodied AI:** The successful transfer to real-world robot tasks and the RoboDojo benchmark demonstrates HDR's potential for physical interaction and embodied intelligence. Low-latency, robust reasoning is critical for robot control, planning, and learning from visual observations. 4. **New Benchmarking Standard:** The introduced level-stratified multi-step video reasoning benchmark provides a valuable tool for future research, encouraging the development of models that prioritize logical consistency over mere visual plausibility. 5. **Architectural Innovation:** The concept of structured multi-scale latent planning and entropy-matched denoising could inspire new architectural designs in other generative domains beyond video, where balancing efficiency and global coherence is important. This paper introduces HDR, a novel framework for multi-step visual reasoning that integrates hierarchical latents and sparse attention into causal video generation. The work effectively bridges the gap between efficient streaming and robust reasoning, demonstrating significant improvements on a new, comprehensive benchmark and showing strong transferability to real-world robot tasks, thus advancing the capabilities of video models as world models for complex, long-horizon planning.
We study aggregation of statistical evidence under unknown and potentially complex dependence using group-invariance. Building on permutation-based constructions that treat transformed datasets as exchangeable units, we aggregate evidence across statistics for each transformed dataset and calibrate the resulting aggregates across transformations. We develop a finite-sample power and adaptivity theory for this framework, together with extensions to sequential and data-dependent aggregation that preserve validity. For single-batch aggregation, which uses one collection of transformed datasets for both standardization and calibration, we show that the critical values uniformly improve on deterministic calibrations valid under arbitrary dependence, including Bonferroni correction, while adapting to the unknown dependence structure. We also introduce a sequential alpha-spending version that permits early rejection when evidence is strong, and a two-batch extension that separates standardization from calibration to accommodate learned aggregation rules and reduce computation. Applications to adaptive nonparametric testing and conformal prediction illustrate how these results sharpen existing aggregation methods.
Primary: University of Cambridge
All Institutions: University of Cambridge, University College London, KAIST
This paper provides a rigorous and novel theoretical framework for aggregating statistical evidence under group-invariance, introducing Single-Batch and Two-Batch aggregation methods that guarantee finite-sample validity while adapting to unknown dependence structures, thereby offering strictly improved power over worst-case calibrations like Bonferroni and resolving finite-sample size control issues in Monte Carlo approximations.
The paper presents a rigorous theoretical framework for aggregating statistical evidence under group-invariance. The core methodological contribution is the formalization of "Single-Batch" (SB) aggregation, which uses a single set of transformed datasets for both standardization and calibration, and the "Two-Batch" (TB) extension, which separates these roles. The authors prove finite-sample validity for these methods under arbitrary dependence structures, a significant improvement over Monte Carlo approximations that may fail to control Type I error in finite samples. The introduction of sequential alpha-spending for ordered statistics and the analysis of adaptivity to dependence structures (perfect rank alignment) are theoretically sound and novel contributions to the theory of permutation tests and p-value merging.
The paper includes applications to adaptive nonparametric testing and conformal prediction. While the text provided is truncated, the abstract and introduction indicate that the theoretical results are illustrated through these applications. The theoretical proofs are extensive, covering validity, power dominance over Bonferroni/deterministic calibrations, and asymptotic adaptivity. The experimental section likely demonstrates empirical performance gains, particularly in conformal prediction where the TB method avoids computational bottlenecks. The rigorous theoretical backing provides strong indirect evidence of utility.
The paper provides detailed algorithms (SB, SeqSB, TB) and mathematical formulations. The theoretical results are clearly stated with proofs (deferred to supplementary material, which is standard for such theoretical work). The reliance on group-invariance makes the methods reproducible given the data and the group structure.
The methods are restricted to settings where group-invariance (exchangeability) holds under the null hypothesis. This limits applicability to observational studies without such symmetries. The "Two-Batch" method requires splitting data, which can reduce power if the sample size is small. The theoretical gains are asymptotic or finite-sample bounds; practical performance depends on the specific dependence structure and the choice of merging function.
This work has significant implications for multiple testing, conformal prediction, and distribution-free inference in machine learning. By providing valid, less conservative aggregation methods under exchangeability, it enables more powerful hypothesis testing and tighter prediction intervals in settings like genetic association studies, image analysis, and any domain where permutation tests are applicable. It bridges the gap between theoretical validity and practical power in dependent data settings. This paper provides a rigorous and novel theoretical framework for aggregating statistical evidence under group-invariance, introducing Single-Batch and Two-Batch aggregation methods that guarantee finite-sample validity while adapting to unknown dependence structures, thereby offering strictly improved power over worst-case calibrations like Bonferroni and resolving finite-sample size control issues in Monte Carlo approximations.
Scaling executable agent training data for LLM post-training is bottlenecked by substrate-bound methods that tie task generation to predefined tools, repositories, or skill graphs: expanding coverage requires manual substrate engineering, each new domain demands a bespoke pipeline, and the resulting task distributions often reflect substrate biases rather than real-world demand. We introduce NexForge, a requirement-driven framework that takes high-level capability requirements as input and synthesizes diverse, executable agent tasks and expert trajectories for SFT. NexForge first investigates real-world demand to construct representative scenarios and task profiles, then performs distribution-aware compilation to generate task directives. For each directive, NexForge automatically retrieves or constructs the required files, dependencies, and runtime configurations, and finally synthesizes expert rollouts and produces training trajectories. Without domain-specific infrastructure, NexForge produces 3.6K terminal and 2K office tasks, improving Qwen3.5-35B-A3B Base from 22.5\% to 52.0\% on Terminal-Bench 2.0 and from 813 to 1338 Elo on GDPval; scaling further to 43.2K terminal tasks yields 58.4\%, on par with Claude Opus 4.6 equipped with Claude Code. Scaled further, NexForge-synthesized data contributes to the training of Nex-N2, a family of publicly available agent models that lift Qwen3.5-35B-A3B to 75.3\% on Terminal-Bench 2.1 and to 1585 Elo on GDPval -- achieving state-of-the-art open-source performance and surpassing several frontier proprietary systems. Nex-N2 models are available at https://nex.sii.edu.cn/.
Primary: SII (Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences)
All Institutions: SII (Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences)
NexForge presents a compelling and effective pipeline for scaling agent training data through requirement-driven synthesis, demonstrating that high-quality, diverse task generation can significantly boost LLM agent performance, achieving state-of-the-art open-source results on key benchmarks.
The paper introduces NexForge, a framework designed to address the data bottleneck in training LLM-based agents. The core innovation lies in shifting from "substrate-bound" task generation (which relies on predefined tools or codebases) to a "requirement-driven" approach. The methodology involves three key stages: 1) Analyzing real-world demand to create representative scenarios and task profiles; 2) Distribution-aware compilation to generate high-level task directives; and 3) Automatic synthesis of executable environments (files, dependencies, runtime configs) and expert rollouts for Supervised Fine-Tuning (SFT). This approach aims to reduce manual engineering and mitigate substrate biases. The method is technically sound, leveraging existing LLM capabilities for code generation and environment setup, but the novelty is incremental rather than foundational. It represents a sophisticated engineering pipeline rather than a new algorithmic breakthrough.
The experimental section demonstrates significant empirical improvements. Using Qwen3.5-35B-A3B as the base, the authors show a jump from 22.5% to 52.0% on Terminal-Bench 2.0 and from 813 to 1338 Elo on GDPval with 3.6K terminal and 2K office tasks. Scaling to 43.2K terminal tasks pushes performance to 58.4%, which is comparable to Claude Opus 4.6 with Claude Code. Furthermore, the synthesized data is used to train "Nex-N2," achieving state-of-the-art open-source results (75.3% on Terminal-Bench 2.1, 1585 Elo on GDPval). The results are impressive and suggest that high-quality, diverse, requirement-driven data is a critical lever for agent performance. The evaluation is rigorous, covering multiple benchmarks and comparing against strong proprietary baselines.
The paper provides a project URL (https://nex.sii.edu.cn/) which likely contains code and model weights. The description of the pipeline (requirement analysis -> directive compilation -> environment synthesis -> rollout) is detailed enough to be reproducible by a team with sufficient resources. However, the "expert rollouts" likely rely on a strong teacher model or human-in-the-loop, which can introduce variability. The specific "distribution-aware compilation" algorithm is not fully detailed in the abstract, so full reproducibility depends on the completeness of the main text and code release.
The paper does not explicitly discuss the cost of generating 43.2K high-quality tasks, which can be significant. The reliance on a "requirement-driven" approach assumes that high-level requirements can be effectively mapped to executable tasks, which may fail in domains with ambiguous or complex implicit constraints. Additionally, the "substrate biases" argument, while valid, might be overstated if the underlying LLMs themselves have biases in their training data that NexForge cannot correct. The evaluation is primarily on coding/terminal tasks; generalization to other agent domains (e.g., web browsing, multi-modal reasoning) is not demonstrated.
This work has significant implications for the democratization of capable AI agents. By providing a scalable method for generating high-quality training data, it lowers the barrier to entry for developing specialized agents. The release of Nex-N2 models contributes to the open-source ecosystem. However, the potential for misuse (e.g., generating malicious code or automating cyberattacks) is a concern that should be addressed in the broader impact statement. The success of such frameworks may accelerate the arms race in agent capabilities, raising safety and alignment challenges. NexForge presents a compelling and effective pipeline for scaling agent training data through requirement-driven synthesis, demonstrating that high-quality, diverse task generation can significantly boost LLM agent performance, achieving state-of-the-art open-source results on key benchmarks.
Summation error depends on partial-sum order, which standard worst-case bounds omit. To capture this dependence, we derive an exact mean-square error (MSE) recurrence for a binary reduction tree T under conditionally unbiased rounding. With unit roundoff u, the constant-nu model sets the local variance at pre-rounding value x to nu u^2 x^2. Its leading tree-dependent cost for the input vector p is p^T K_T p, where the common-ancestor kernel K_T counts the internal ancestors shared by each pair of leaves. For i.i.d. inputs of mean mu and variance tau^2, this expected cost is tau^2 Lambda_1(T) + mu^2 Lambda_2(T), where Lambda_1 is total leaf depth and Lambda_2 sums squared internal-subtree sizes; Lambda_1 governs centered inputs, while Lambda_2 captures nonzero means. We use these statistics to characterize optimal tree topologies and schedules. Balanced and sequential trees attain the centered extrema. For k inputs, optimal two-stage sequential blocking yields root-mean-square (RMS) error scaling as k^{3/4}. For fixed-stage hierarchies, geometric schedules are optimal for centered inputs, whereas the optimal noncentered stage exponents halve successively. For independent centered inputs with unequal variances, Huffman coding minimizes variance-weighted depth over free leaf assignments. We extend the kernel to matrix multiplication through operand Gram matrices. We then test the approximation under round-to-nearest using exact residuals. Across binary64, binary32, and software-emulated binary16 and bfloat16, the model recovers the ordering among tree topologies; K_T tracks AR(1) partial-sum costs. For GEMM, independently calibrated predictions differ from measurements by at most 3% on the tested grid. A reduction tree extracted from an array library predicts the measured RMS scaling. However, stagnation and bias in positive low-precision sums limit the model's applicability.
Primary: Oak Ridge National Laboratory
All Institutions: Oak Ridge National Laboratory
This paper provides a rigorous second-moment theory for floating-point reduction trees, deriving exact error recurrences and optimal topologies that significantly advance the understanding of numerical stability in parallel reductions, with direct applications to improving the accuracy of HPC and ML libraries.
The paper presents a rigorous theoretical framework for analyzing floating-point reduction errors, specifically focusing on the variance of partial sums in reduction trees. The core methodological contribution is the derivation of an exact mean-square error (MSE) recurrence relation for binary reduction trees under the conditionally unbiased rounding model. The authors introduce the "common-ancestor kernel" $K_T$, which quantifies the structural impact of tree topology on error propagation. They decompose the expected cost into terms dependent on input variance ($\Lambda_1$) and mean ($\Lambda_2$), allowing for the characterization of optimal tree topologies (e.g., balanced vs. sequential, Huffman coding for unequal variances). The methodology extends naturally to matrix multiplication via Gram matrices, providing a unified view of reduction error in linear algebra operations. The approach is mathematically sound, leveraging statistical properties of floating-point arithmetic rather than worst-case bounds, which offers a more realistic model for typical workloads.
The authors validate their theoretical model through extensive experiments across multiple precision formats (binary64, binary32, software-emulated binary16, and bfloat16). They demonstrate that the model accurately predicts the ordering of RMS error among different tree topologies and tracks the costs of autoregressive (AR(1)) processes. For General Matrix Multiplication (GEMM), the model's predictions differ from measurements by at most 3% on the tested grid. They also test a reduction tree extracted from an actual array library, confirming the model's predictive power for real-world implementations. The experiments are well-designed, covering both synthetic i.i.d. inputs and structured data, and effectively bridge the gap between theoretical bounds and empirical behavior.
The paper provides detailed mathematical derivations and specifies the rounding models and input distributions used in experiments. The mention of "software-emulated" formats suggests that the authors have implemented or utilized existing tools for lower-precision simulation, which aids reproducibility. However, the paper does not explicitly provide a link to the source code or specific software versions used for the simulations in the abstract or main text provided. Given the institutional context (ORNL) and the nature of the work, code is likely available or reproducible, but explicit URLs are missing from the provided text.
The authors explicitly acknowledge limitations, noting that the model's applicability is limited by stagnation and bias in positive low-precision sums. This suggests that the conditionally unbiased rounding assumption may break down in specific edge cases, particularly with low-precision formats like bfloat16 or binary16 where dynamic range and precision constraints are tighter. The model is primarily statistical (expectation/variance) and may not capture worst-case scenarios or specific pathological inputs that trigger catastrophic cancellation or overflow in ways not captured by the second-moment analysis.
This work has significant implications for the design of high-performance computing (HPC) libraries and machine learning frameworks that rely heavily on parallel reductions (e.g., dot products, sums, matrix multiplications). By providing a theory for optimal tree topologies based on input statistics, it enables the development of adaptive algorithms that minimize numerical error without sacrificing performance. This is particularly relevant for mixed-precision training and inference, where understanding and controlling error propagation is critical. The insights could lead to more robust and accurate numerical libraries for deep learning and scientific computing. This paper provides a rigorous second-moment theory for floating-point reduction trees, deriving exact error recurrences and optimal topologies that significantly advance the understanding of numerical stability in parallel reductions, with direct applications to improving the accuracy of HPC and ML libraries.
Reinforcement learning with verifiable rewards (RLVR) has substantially improved the reasoning capabilities of large language models on tasks such as mathematical reasoning and code generation. However, most RLVR methods assign a scalar outcome reward to an entire trajectory, resulting in sparse supervision and limited token-level credit assignment. On-policy distillation (OPD) provides denser supervision by distilling token-level distributions from a stronger teacher model, but requires an additional teacher and typically assumes a shared vocabulary. On-policy self-distillation (OPSD) removes this dependency by conditioning the same model on privileged information to construct a teacher policy. However, directly matching the teacher distribution may cause information leakage and unstable optimization. RLSD avoids direct matching by using the teacher signal only to modulate update magnitudes, but it cannot provide an explicit correction direction when the sampled reasoning fails. To address this tradeoff, we introduce H^{2}SD, a hybrid hindsight self distillation framework that uses the teacher differently according to trajectory correctness. For successful trajectories, the teacher receives the student response confirmed as correct together with a rephrasing instruction, and its probabilities on the original response tokens are used to modulate update magnitudes without changing the direction determined by the reward. For failed trajectories, we condition the teacher on a reference hint containing key reasoning steps and a verified answer, and minimize the reverse KL divergence from the student to the teacher. Experiments on multiple challenging reasoning benchmarks show that H^2SD consistently outperforms representative RLVR, OPSD, and RLSD baselines while maintaining stable optimization and favorable generation efficiency.
Primary: Harbin Institute of Technology
All Institutions: Harbin Institute of Technology, Beijing Academy of Artificial Intelligence
H^2SD presents a nuanced approach to self-distillation in RLVR by differentiating between successful and failed trajectories, offering a potential path to more stable and effective training for reasoning models. The hybrid strategy of magnitude modulation for successes and reverse KL correction for failures is a logical and promising direction, though its practical utility hinges on the efficiency of the hint generation and the computational overhead it introduces.
The paper proposes H^2SD, a hybrid hindsight self-distillation framework designed to address the sparse supervision and credit assignment issues in Reinforcement Learning with Verifiable Rewards (RLVR). The core innovation lies in the conditional use of a teacher policy (the model itself, conditioned on privileged information) based on trajectory correctness. For successful trajectories, the teacher's probabilities modulate update magnitudes (similar to RLSD) to stabilize training without altering the reward-driven direction. For failed trajectories, the method minimizes reverse KL divergence from a teacher conditioned on a reference hint (key reasoning steps and verified answer), providing an explicit correction direction. This hybrid approach attempts to balance the stability of policy gradient methods with the dense supervision of distillation. The methodology is theoretically sound, drawing on established concepts in hindsight experience replay and self-distillation, but the specific combination and the mechanism for generating the "reference hint" are the key technical contributions.
The authors evaluate H^2SD on multiple challenging reasoning benchmarks, comparing it against RLVR, OPSD, and RLSD baselines. The results indicate consistent outperformance in terms of accuracy and stable optimization. The paper demonstrates that the hybrid approach mitigates the instability of direct matching and the lack of correction direction in pure RL methods. However, as an arXiv preprint with 0 citations, the breadth of the experimental validation (e.g., ablation studies on the hint generation quality, sensitivity to hyperparameters, and computational overhead) is critical. The claim of "favorable generation efficiency" needs scrutiny, as adding a teacher pass (even if self-distilled) incurs inference costs during training. The results appear promising but require independent verification to confirm generalizability across different model sizes and reasoning domains.
The paper provides a clear description of the methodology, including the loss functions and the conditioning mechanism. The use of standard RLVR setups suggests that the codebase could be built upon existing frameworks like OpenRL or TRL. However, the specific implementation details of the "reference hint" generation (whether it is heuristic, model-generated, or retrieved) are crucial for reproducibility. If the hint generation is complex or relies on external tools not fully detailed, reproducibility may be hindered. The authors should provide code to ensure the community can replicate the results.
A significant limitation is the reliance on the quality of the "reference hint." If the hint is not informative or accurate, the reverse KL minimization on failed trajectories could lead to poor learning or mode collapse. Additionally, the computational cost of running the teacher model (even if it is the student model with different conditioning) during every training step increases the training time and memory footprint compared to standard RLVR. The paper does not explicitly quantify this overhead. Furthermore, the assumption that the model can effectively condition on "privileged information" (the correct answer/hint) during training but not during inference requires careful handling to avoid distribution shift issues.
This work contributes to the broader goal of improving the reasoning capabilities of LLMs, which has significant implications for automated theorem proving, code generation, and complex decision-making systems. By making RLVR more stable and sample-efficient, it lowers the barrier to training capable reasoning models. However, the increased computational cost during training could limit accessibility for smaller research groups. The focus on mathematical and code reasoning also highlights the potential for these methods to be applied to other structured reasoning tasks. H^2SD presents a nuanced approach to self-distillation in RLVR by differentiating between successful and failed trajectories, offering a potential path to more stable and effective training for reasoning models. The hybrid strategy of magnitude modulation for successes and reverse KL correction for failures is a logical and promising direction, though its practical utility hinges on the efficiency of the hint generation and the computational overhead it introduces.
Pruning long context for coding agents has been a vital technology for efficient context management. While existing context pruning methods such as SWE-Pruner realize this by attaching a separate code classifier, we find the agent itself encodes internal representations indicating the relevance of code context when reading tool output. Based on this finding, we propose SWE-Pruner Pro, which prunes tool outputs directly inside the agent. Concretely, a small head turns the agent's own internal representations into a keep-or-prune label for each line, with a length-aware embedding keyed to each tool output's line count. Across two open-weight backbones and four multi-turn benchmarks, SWE-Pruner Pro saves up to 39% of prompt and completion tokens while preserving task quality, with bounded inference overhead. Notably, on MiMo-V2-Flash SWE-Pruner Pro additionally raises the SWE-Bench Verified resolve rate by +3.8% and the long-context Oolong accuracy by +2.2 points.
Primary: Shanghai Jiao Tong University
All Institutions: Shanghai Jiao Tong University
SWE-Pruner Pro demonstrates that coding agents' internal representations encode sufficient line-level relevance information to prune tool outputs effectively, achieving significant token savings and improved efficiency without retraining the backbone, marking a practical step forward in scalable agentic systems.
The paper proposes SWE-Pruner Pro, a method for pruning long context in coding agents by extracting a lightweight classification head from the agent's own internal hidden states. The core insight is that the backbone model already encodes line-level relevance information during the forward pass of tool outputs, eliminating the need for a separate scoring model or explicit goal-hint queries used in prior work (e.g., SWE-Pruner). The methodology involves a length-aware embedding and a per-sample balanced focal loss to handle class imbalance and length-dependent error costs. The approach is technically sound, leveraging existing inference infrastructure (SGLang) with specific patches to expose hidden states. The design is pragmatic, focusing on efficiency gains without retraining the backbone.
The evaluation is comprehensive, covering two large open-weight backbones (MiMo-V2-Flash, Qwen3-Coder-Next) and four benchmarks (SWE-Bench Verified, SWE-QA, SWE-QA-Pro, Oolong). The results demonstrate up to 39% token savings while preserving or slightly improving task quality (e.g., +3.8% resolve rate on SWE-Bench Verified). The inclusion of an ablation study on loss functions and length-aware embeddings adds rigor. The comparison against seven baselines, including strong prior work, provides a solid empirical foundation. The latency analysis, including in-engine colocation, further strengthens the practical value of the work.
The paper provides detailed implementation details, including the architecture of the pruning head, training data sources (publicly released datasets), and specific SGLang patches required for hidden state extraction. The training data distribution and labeling protocol are described. However, the code for the SGLang patches and the specific pruning head implementation are not explicitly linked in the text (though implied to be available or part of the project). The reliance on specific versions of SGLang and the need for custom patches might pose minor reproducibility hurdles for users not familiar with the inference engine's internals.
The method is currently limited to open-weight models that expose hidden states. The evaluation is primarily focused on Python and CLI tasks, with limited coverage of other programming languages. The pruning is applied to tool outputs, so it does not address pruning of the agent's own reasoning or history, which may also contain redundant information. The performance gain on SWE-Bench Verified is notable but the absolute resolve rate is still constrained by the backbone's capabilities.
This work significantly advances the field of efficient LLM inference, particularly for agentic workflows. By demonstrating that internal representations contain sufficient signal for pruning, it reduces the computational overhead and latency associated with context management. This can lead to more cost-effective and scalable deployment of coding agents, enabling longer and more complex interactions. The approach could be generalized to other domains where agents interact with long textual environments. SWE-Pruner Pro demonstrates that coding agents' internal representations encode sufficient line-level relevance information to prune tool outputs effectively, achieving significant token savings and improved efficiency without retraining the backbone, marking a practical step forward in scalable agentic systems.
We present RynnBrain 1.1, a family of embodied foundation models spanning 2B, 9B, and 122B-A10B scales. Trained with a unified spatio-temporal and physically grounded framework, RynnBrain 1.1 supports embodied perception, spatial reasoning, localization, and planning. Compared with RynnBrain 1.0, it further introduces contact-point prediction across the model family and native 3D grounding for the 2B and 9B models, yielding representations and outputs that are more directly aligned with robot manipulation. We also develop RynnBrain-VLA with a unified cross-embodiment action space and embodiment-specific masking, and deploy it on Unitree G1, Astribot-S1, and Tianji-Wuji. RynnBrain 1.1 achieves strong results on embodied cognition, localization, and 3D grounding, with the 122B-A10B model outperforming all evaluated proprietary and open-source models on VSI-Bench, MMSI, and RefSpatial-Bench. Real-robot experiments show that RynnBrain-initialized policies outperform Qwen-based and representative generalist VLAs, while joint multi-task and multi-embodiment training improves process scores and success rates over per-task training.
Primary: Rynn AI
All Institutions: Rynn AI
RynnBrain 1.1 presents a significant advancement in embodied foundation models by introducing native 3D grounding and contact-point prediction within a unified spatio-temporal framework, demonstrating strong generalization across multiple robotic platforms and outperforming existing models on key embodied cognition benchmarks.
The paper introduces RynnBrain 1.1, a family of embodied foundation models (2B, 9B, 122B-A10B). The core methodological contribution lies in a unified spatio-temporal and physically grounded pretraining framework. Key technical innovations include the introduction of contact-point prediction across the model family and native 3D grounding for the smaller models (2B and 9B). The authors also propose RynnBrain-VLA, which utilizes a unified cross-embodiment action space and embodiment-specific masking to handle diverse robotic hardware. The architecture appears to scale from dense to Mixture-of-Experts (MoE) models, aiming to balance capability with inference efficiency. The approach integrates perception, spatial reasoning, and planning into a single unified representation, which is a significant step towards generalist embodied agents.
The evaluation covers embodied cognition, localization, and 3D grounding benchmarks (VSI-Bench, MMSI, RefSpatial-Bench). The 122B-A10B model claims to outperform all evaluated proprietary and open-source models on these benchmarks. Real-robot experiments are conducted on three distinct platforms: Unitree G1, Astribot-S1, and Tianji-Wuji. The results indicate that RynnBrain-initialized policies outperform Qwen-based and other generalist VLAs. The paper also highlights the benefits of joint multi-task and multi-embodiment training over per-task training, showing improvements in process scores and success rates. The breadth of evaluation across simulation and real-world hardware is a strong point, demonstrating generalization capabilities.
The paper provides details on model scales and training frameworks. However, as an arXiv preprint without an accompanying code release mentioned in the text, reproducibility is currently limited to the described methodology. The use of specific hardware (Unitree, Astribot, Tianji) for real-world evaluation adds a layer of complexity for independent verification, requiring access to similar robotic platforms. The mention of "RynnBrain" suggests a proprietary model family, which may limit full open-source reproducibility of the weights.
The paper focuses heavily on the capabilities of the 122B-A10B model for benchmark leadership, but the performance of the smaller 2B and 9B models, while improved, may not match the state-of-the-art in every metric. The reliance on embodiment-specific masking and unified action spaces requires careful calibration for each new robot type, which might limit plug-and-play generalization to unseen embodiments without further tuning. The computational cost of training and deploying the 122B model is significant, potentially limiting accessibility for smaller research groups.
This work contributes to the development of generalist embodied AI, which has profound implications for robotics, automation, and human-robot interaction. By providing models that can reason about 3D space and physical interactions, it paves the way for more capable and autonomous robots in unstructured environments. The emphasis on generalization across embodiments could accelerate the deployment of robots in various industries. However, the increased autonomy of robots also raises safety and ethical considerations regarding control and reliability in physical spaces. RynnBrain 1.1 presents a significant advancement in embodied foundation models by introducing native 3D grounding and contact-point prediction within a unified spatio-temporal framework, demonstrating strong generalization across multiple robotic platforms and outperforming existing models on key embodied cognition benchmarks.
Video multimodal large language models (MLLMs) can describe what happens in a video, but rarely identify when the supporting evidence occurs. We study generalist video temporal grounding, in which one model predicts a variable-cardinality set of evidence intervals across video lengths, domains, query forms, and viewpoints. Existing training strategies are misaligned with this set-valued task: long-video labels often rely on brittle one-pass annotation, while reinforcement-learning rewards either fail to distinguish non-overlapping predictions or require fragile segment matching. TimeLens2 treats temporal evidence as an interval set throughout supervision and optimization. TimeLens2-93K constructs reliable multi-span supervision through caption-derived proposals, independent localization, cross-agent consensus, semantic verification, and boundary refinement. Our temporal Wasserstein reward computes exact one-dimensional \(W_1\) between uniform distributions over merged interval supports, providing dense, matching-free feedback under unequal cardinalities and equivalent fragmentation; temporal IoU complements it with precise-overlap feedback. Across seven benchmarks, TimeLens2-2B outperforms all size-matched baselines on every benchmark, while the 4B and 8B variants achieve state-of-the-art performance, surpassing open-source models with up to 397B parameters. The 2B, 4B, and 8B variants improve over their Qwen3-VL backbones by 14.2, 13.0, and 18.1 mIoU points, respectively.
Primary: Nanjing University
All Institutions: Nanjing University
TimeLens2 presents a significant advancement in video temporal grounding by introducing a temporal Wasserstein reward and a robust multi-span supervision pipeline, achieving state-of-the-art performance with efficient model sizes. The technical contribution is solid, addressing a key limitation in current MLLM training for set-valued video tasks, and the empirical results demonstrate clear improvements over existing baselines, warranting a high score for its potential impact on the field of video understanding.
The paper proposes TimeLens2, a generalist video temporal grounding (VTG) framework that leverages Multimodal Large Language Models (MLLMs). The core methodological innovation lies in addressing the misalignment between existing training strategies and the set-valued nature of VTG (variable number of intervals). Specifically, it introduces a "temporal Wasserstein reward" that computes the $W_1$ distance between uniform distributions over merged interval supports. This provides dense, matching-free feedback that handles unequal cardinalities and fragmentation issues better than standard IoU or RL rewards. The training data construction (TimeLens2-93K) uses a multi-stage pipeline involving caption-derived proposals, independent localization, cross-agent consensus, and boundary refinement to create reliable multi-span supervision. This approach is technically sound and addresses a specific, non-trivial pain point in VTG: the evaluation and training of models that predict variable-length sets of intervals.
The authors evaluate TimeLens2 across seven benchmarks. The results show that the 2B variant outperforms all size-matched baselines, and the 4B/8B variants achieve state-of-the-art performance, surpassing much larger open-source models (up to 397B parameters). The improvements over the Qwen3-VL backbones are significant (14.2 to 18.1 mIoU points). The experimental setup appears rigorous, covering multiple domains and query forms. The comparison against large parameter models highlights the efficiency and effectiveness of the proposed method. However, the reliance on "seven benchmarks" suggests a broad but potentially shallow evaluation compared to papers that provide deep ablation studies on specific failure modes or domain shifts.
The paper mentions the construction of a 93K dataset (TimeLens2-93K) and details the pipeline for its creation. The use of standard MLLM backbones (Qwen3-VL) and clear mathematical definitions for the Wasserstein reward enhances reproducibility. The lack of a provided code repository in the metadata is a minor negative, but the methodological description seems sufficient for replication by competent researchers.
The paper does not explicitly discuss the computational cost of the proposed Wasserstein reward calculation during training or inference compared to standard IoU. The "cross-agent consensus" step in data generation might introduce biases or errors if the underlying models are not robust. Furthermore, while it claims "generalist" capabilities, the performance on highly specialized or out-of-distribution domains (e.g., medical video, scientific diagrams) is not detailed, which is a common limitation for generalist models. The reliance on Qwen3-VL (which appears to be a very recent or future-dated model given the ICLR 2026 venue) means the results are tightly coupled to that specific backbone's capabilities.
This work contributes to the advancement of video understanding systems, specifically in applications requiring precise temporal localization (e.g., video retrieval, summarization, interactive video analysis). By improving the efficiency and accuracy of MLLMs in VTG, it enables more capable and accessible video AI tools. The focus on set-valued tasks also has broader implications for other sequential prediction tasks in vision-language models. TimeLens2 presents a significant advancement in video temporal grounding by introducing a temporal Wasserstein reward and a robust multi-span supervision pipeline, achieving state-of-the-art performance with efficient model sizes. The technical contribution is solid, addressing a key limitation in current MLLM training for set-valued video tasks, and the empirical results demonstrate clear improvements over existing baselines, warranting a high score for its potential impact on the field of video understanding.
This paper introduces EvolvingWorld, a framework and benchmark for character and world co-evolution in interactive literary worlds. Existing systems either treat interactive literary simulation as static persona imitation or isolated scene generation, failing to capture how characters and worlds evolve together over time. To address this, EvolvingWorld models literary simulation as a long-horizon process where characters interact, scenes progress, and character and world states are persistently updated. Unlike prior systems relying on fixed schemas, EvolvingWorld adopts an open-schema framework to support simulation across diverse literary worlds. The framework consists of two coupled modules: a Character Agent for multi-character role-play and persistent profile evolution, and an LLM-based World Model for global and location/entity-level state maintenance and scene progression. Based on this architecture, we formulate 7 trainable tasks for scene initialization, interaction generation, and state update. We construct a dataset from 57 books, producing 138,596 supervised training samples and 222 snapshots for testing. Furthermore, we introduce a trajectory-level LLM-as-Judge evaluation protocol spanning 10 dimensions and 20 metrics. Experiments show that EvolvingWorld can improve long-horizon simulation by effectively maintaining persistent, coherent character and world development.
Primary: Hong Kong University of Science and Technology
All Institutions: Hong Kong University of Science and Technology, Huazhong University of Science and Technology
EvolvingWorld introduces a novel open-schema framework for co-evolving character and world states in interactive literary simulations, providing a significant methodological and benchmarking contribution to the field of long-horizon multi-agent systems.
The paper proposes "EvolvingWorld," a framework for long-horizon interactive literary simulation. The core methodological contribution is the decomposition of this simulation into two coupled modules: an open-schema Character Agent and an LLM-based World Model. The "open-schema" aspect is notable, as it allows the system to infer relevant character and world dimensions dynamically rather than relying on fixed ontologies. The introduction of "hidden trackers" for multi-timescale profile evolution (distinguishing between rapid mood shifts and slow personality changes) is a thoughtful architectural detail. The simulation pipeline is rigorously defined with seven trainable tasks (scene casting, location scenario, motivation update, next character, interaction generation, world update, character update). This structured decomposition transforms a vague generative task into a supervised learning problem, which is a significant methodological step for making long-horizon agents trainable and evaluable.
The authors construct a substantial dataset from 57 public-domain books, creating over 138k training samples. They introduce a comprehensive evaluation protocol with 20 metrics across 10 dimensions, addressing the lack of standardized benchmarks for long-horizon state evolution. The experiments compare fine-tuned open-source models (Qwen, Llama) against strong closed-source baselines (GPT-4o, Claude, Gemini) and prior role-play baselines (CoSER, Crab, BookWorld). The results demonstrate that supervised fine-tuning on the EvolvingWorld data significantly improves performance on character consistency, evolution quality, and world state maintenance compared to instruction-tuned baselines and prior role-play specific models. The comparison with BookWorld is particularly strong, highlighting the advantages of open-schema and entity-level state tracking. The inclusion of both in-domain and out-of-domain tests adds robustness to the claims.
The paper provides a GitHub repository link. The dataset construction process is described in detail, including the use of LLMs for extraction and the specific handling of chronological narratives. The training details (LoRA, batch size, learning rate) are provided in the appendix. The evaluation metrics and LLM-as-Judge protocols are well-defined. The use of public-domain books ensures the data is accessible. However, the reliance on LLM-as-Judge for evaluation introduces some subjectivity, though the authors attempt to mitigate this with multi-judge setups and human evaluation correlations.
The paper acknowledges several limitations. First, the world model assumes a single objective state, ignoring subjective perceptions or imperfect memories of characters, which is a significant simplification for literary realism. Second, the world representation is constrained by context length, tracking only "important" entities. Third, the dataset is limited to public-domain classic books, which may not generalize well to modern genres or user-created worlds, although the open-schema design suggests potential for adaptation. The evaluation relies heavily on LLM judges, which can be biased or inconsistent.
EvolvingWorld provides a foundational framework for creating more persistent and coherent interactive agents, with applications in gaming, interactive storytelling, and educational simulations. By shifting the focus from static persona imitation to dynamic co-evolution, it pushes the field toward more realistic and engaging long-horizon interactions. The benchmark and dataset will likely serve as a standard for evaluating future long-horizon agent systems. EvolvingWorld introduces a novel open-schema framework for co-evolving character and world states in interactive literary simulations, providing a significant methodological and benchmarking contribution to the field of long-horizon multi-agent systems.
Training API-calling large language model (LLM) agents demands massive amounts of high-quality trajectories. However, collecting such data at scale typically requires fully implemented environments with executable APIs and realistic, pre-populated backend databases, creating a major bottleneck for scalability. To overcome this, we propose an environment-free synthetic data generation approach that leverages LLMs as on-the-fly digital world models. Given only API specifications, our method generates trajectories mimicking interactions between an agent and a stateful environment. Specifically, an LLM first generates diverse tasks solvable with the provided APIs. A teacher agent then iteratively solves each task while an LLM simulator generates coherent synthetic API responses conditioned on the task context and simulation history. Finally, an LLM judge filters the trajectories to ensure the quality of the resulting dataset. We evaluate our approach on the challenging AppWorld and OfficeBench benchmarks, which include both information-retrieval and state-changing tasks. Fine-tuning models on our synthetic data yields significant performance gains, demonstrating that effective supervision for API-calling agents can be generated without any executable environment. Our results establish LLM-based API simulation as a practical, scalable solution for training agents across diverse API ecosystems.
Primary: Unknown
All Institutions: Unknown
This paper presents a practical and scalable method for generating synthetic training data for API-calling agents by leveraging LLMs as dynamic world models, effectively bypassing the need for executable environments and demonstrating significant performance improvements on standard benchmarks.
The paper proposes a pipeline for generating synthetic training data for API-calling agents without requiring an executable environment. The core methodology involves three stages: (1) Task Generation: An LLM generates diverse tasks based on API specifications. (2) Trajectory Simulation: A "teacher" agent attempts to solve these tasks, while a separate "simulator" LLM generates the API responses and state transitions on the fly, conditioned on the context. (3) Filtering: An LLM judge filters the resulting trajectories for quality. The approach leverages the reasoning capabilities of LLMs to model the environment dynamics implicitly, bypassing the need for complex backend implementations. While the concept of using LLMs as world models is not entirely new, applying it specifically to the high-dimensional, state-dependent space of API interactions for data generation is a pragmatic and novel application. The separation of the teacher (reasoning) and simulator (state modeling) roles is a key methodological choice that simplifies the training objective for the student agent.
The authors evaluate their method on AppWorld and OfficeBench, two challenging benchmarks involving information retrieval and state-changing tasks. The primary metric is the performance of models fine-tuned on the generated synthetic data compared to baselines (likely zero-shot or few-shot LLMs, and potentially models fine-tuned on real data if available, though the abstract emphasizes the "environment-free" aspect). The results show "significant performance gains," suggesting that the synthetic data captures useful patterns for agent behavior. However, the abstract does not specify the magnitude of these gains relative to state-of-the-art models trained on real data. The evaluation focuses on the utility of the generated data rather than the fidelity of the simulation itself, which is a valid but limited scope. The use of established benchmarks ensures comparability, but the lack of a real-data baseline in the abstract's summary leaves the absolute ceiling of this approach unclear.
The paper describes a clear, modular pipeline. The use of standard LLMs for task generation, simulation, and filtering suggests that the method is reproducible, provided the specific prompts and model versions are disclosed. The dependency on LLMs for simulation introduces some stochasticity, but the filtering step should mitigate low-quality outputs. The lack of code release (indicated by "none" for Project URL) is a barrier to full reproducibility, but the methodology is described sufficiently to allow for implementation.
The primary limitation is the potential for error propagation. If the simulator generates inconsistent API responses or the teacher agent makes logical errors, the synthetic data will be noisy. The "LLM judge" is intended to filter this, but judges can also be biased or inconsistent. Furthermore, the simulation is only as good as the LLM's understanding of the API semantics; it may hallucinate behaviors that are not possible in the real world. The approach also assumes that the API specifications are sufficient for an LLM to generate coherent interactions, which may not hold for complex, undocumented, or highly dynamic APIs. Finally, the computational cost of generating and filtering large-scale synthetic data using multiple LLM calls is significant.
This work addresses a critical bottleneck in agent development: the scarcity of high-quality training data. By enabling the generation of synthetic data without expensive environments, it lowers the barrier to entry for developing robust API-calling agents. This could accelerate the deployment of agents in diverse domains where building executable environments is difficult or costly. However, it also raises concerns about the quality and safety of training data, as synthetic data may reinforce biases or hallucinations present in the base models. This paper presents a practical and scalable method for generating synthetic training data for API-calling agents by leveraging LLMs as dynamic world models, effectively bypassing the need for executable environments and demonstrating significant performance improvements on standard benchmarks.
We study aggregation of statistical evidence under unknown and potentially complex dependence using group-invariance. Building on permutation-based constructions that treat transformed datasets as exchangeable units, we aggregate evidence across statistics for each transformed dataset and calibrate the resulting aggregates across transformations. We develop a finite-sample power and adaptivity theory for this framework, together with extensions to sequential and data-dependent aggregation that preserve validity. For single-batch aggregation, which uses one collection of transformed datasets for both standardization and calibration, we show that the critical values uniformly improve on deterministic calibrations valid under arbitrary dependence, including Bonferroni correction, while adapting to the unknown dependence structure. We also introduce a sequential alpha-spending version that permits early rejection when evidence is strong, and a two-batch extension that separates standardization from calibration to accommodate learned aggregation rules and reduce computation. Applications to adaptive nonparametric testing and conformal prediction illustrate how these results sharpen existing aggregation methods.
Primary: University of Cambridge
All Institutions: University of Cambridge, University College London, KAIST
This paper provides a rigorous and novel theoretical framework for aggregating statistical evidence under group-invariance, introducing Single-Batch and Two-Batch aggregation methods that guarantee finite-sample validity while adapting to unknown dependence structures, thereby offering strictly improved power over worst-case calibrations like Bonferroni and resolving finite-sample size control issues in Monte Carlo approximations.
The paper presents a rigorous theoretical framework for aggregating statistical evidence under group-invariance. The core methodological contribution is the formalization of "Single-Batch" (SB) aggregation, which uses a single set of transformed datasets for both standardization and calibration, and the "Two-Batch" (TB) extension, which separates these roles. The authors prove finite-sample validity for these methods under arbitrary dependence structures, a significant improvement over Monte Carlo approximations that may fail to control Type I error in finite samples. The introduction of sequential alpha-spending for ordered statistics and the analysis of adaptivity to dependence structures (perfect rank alignment) are theoretically sound and novel contributions to the theory of permutation tests and p-value merging.
The paper includes applications to adaptive nonparametric testing and conformal prediction. While the text provided is truncated, the abstract and introduction indicate that the theoretical results are illustrated through these applications. The theoretical proofs are extensive, covering validity, power dominance over Bonferroni/deterministic calibrations, and asymptotic adaptivity. The experimental section likely demonstrates empirical performance gains, particularly in conformal prediction where the TB method avoids computational bottlenecks. The rigorous theoretical backing provides strong indirect evidence of utility.
The paper provides detailed algorithms (SB, SeqSB, TB) and mathematical formulations. The theoretical results are clearly stated with proofs (deferred to supplementary material, which is standard for such theoretical work). The reliance on group-invariance makes the methods reproducible given the data and the group structure.
The methods are restricted to settings where group-invariance (exchangeability) holds under the null hypothesis. This limits applicability to observational studies without such symmetries. The "Two-Batch" method requires splitting data, which can reduce power if the sample size is small. The theoretical gains are asymptotic or finite-sample bounds; practical performance depends on the specific dependence structure and the choice of merging function.
This work has significant implications for multiple testing, conformal prediction, and distribution-free inference in machine learning. By providing valid, less conservative aggregation methods under exchangeability, it enables more powerful hypothesis testing and tighter prediction intervals in settings like genetic association studies, image analysis, and any domain where permutation tests are applicable. It bridges the gap between theoretical validity and practical power in dependent data settings. This paper provides a rigorous and novel theoretical framework for aggregating statistical evidence under group-invariance, introducing Single-Batch and Two-Batch aggregation methods that guarantee finite-sample validity while adapting to unknown dependence structures, thereby offering strictly improved power over worst-case calibrations like Bonferroni and resolving finite-sample size control issues in Monte Carlo approximations.
Higher-order couplings enhance the expressive power of hypergraph neural networks (HGNNs), but they also intensify representation collapse in deep propagation due to strong multi-way feature mixing. This work investigates hypergraph oversmoothing from a dynamical-systems perspective and develops a reaction--diffusion framework for depth-resistant hypergraph learning. By defining hypergraph gradient and divergence operators, we interpret message passing as an incidence-level diffusion process. The analysis of pure diffusion shows that its continuous semiflow exponentially contracts the null-mode-free component of node representations and drives the Dirichlet energy to zero, revealing hypergraph oversmoothing as an intrinsic transverse-energy dissipation phenomenon. Motivated by this analysis, we propose Hypergraph Neural Reaction--Diffusion (HNRD), which introduces a reaction mechanism acting on the transverse component to compensate diffusion-induced dissipation and stabilize discriminative variations. We establish global well-posedness of the proposed dynamics and prove that the null-mode-free Dirichlet energy remains bounded away from zero. A forward-Euler discretization provides a practical HNRD layer with a stability condition for deep propagation. Experiments on benchmark and synthetic heterophilic hypergraphs demonstrate that HNRD consistently improves over representative hypergraph baselines. Depth, robustness, and efficiency analyses further show that HNRD preserves stable performance and nonzero Dirichlet energy under deep propagation and perturbations. These results provide a principled dynamical framework for designing deep hypergraph architectures that maintain higher-order expressiveness without representation collapse.
Primary: Chinese Academy of Sciences
All Institutions: Shandong University, Chinese Academy of Sciences, University of Chinese Academy of Sciences
This work has significant broader impact for the field of graph and hypergraph neural networks: * **Principled Deep HGNN Design**: It provides a principled, theoretically grounded framework for designing deep HGNNs that inherently mitigate oversmoothing, moving beyond heuristic solutions like residual connections. This can lead to more robust and performant deep hypergraph architectures. * **Deeper Understanding of Oversmoothing**: The dynamical-systems view offers a powerful new lens for understanding oversmoothing in hypergraphs, characterizing it as transverse-energy dissipation. This theoretical insight can guide future research into other graph/hypergraph problems. * **Advancement of Continuous-Time GNNs**: By extending the reaction-diffusion paradigm to hypergraphs with strong theoretical guarantees, the paper further validates the utility of continuous-time models (ODEs/PDEs) in graph machine learning, potentially encouraging more research in this direction. * **Applications in Complex Systems**: Hypergraphs are crucial for modeling multi-way interactions in various domains (social networks, biology, chemistry, recommender systems). A robust deep HGNN model like HNRD can significantly improve performance in these applications, especially where deep propagation is necessary to capture complex dependencies. * **Theoretical Foundation for HGNNs**: The rigorous mathematical analysis contributes to building a stronger theoretical foundation for hypergraph learning, which is essential for the maturity and reliability of the field. This paper introduces a novel dynamical-systems framework to analyze and mitigate oversmoothing in hypergraph neural networks, rigorously proving that pure hypergraph diffusion leads to representation collapse and proposing Hypergraph Neural Reaction–Diffusion (HNRD) which provably stabilizes discriminative variations. The paper's strength lies in its deep theoretical analysis, which precisely characterizes hypergraph oversmoothing as transverse-energy dissipation, and its innovative reaction-diffusion mechanism that explicitly counteracts this collapse by preserving a learnable, nonzero transverse-energy level. This principled approach, combined with comprehensive experimental validation across diverse hypergraph benchmarks and detailed analyses (depth, heterophily, ablation, robustness, visualization), demonstrates HNRD's superior performance and stability in deep and challenging settings. The work significantly advances the theoretical understanding and practical design of deep hypergraph architectures, offering a robust solution to a critical problem and paving the way for more principled continuous-time models in higher-order graph learning.
The paper presents a highly rigorous and principled approach to understanding and mitigating oversmoothing in Hypergraph Neural Networks (HGNNs) through a dynamical-systems lens. The methodology is built upon defining hypergraph gradient and divergence operators at the incidence level, which allows for formulating message passing as a learnable incidence-level diffusion equation. This continuous-time formulation is a significant strength, enabling the use of semiflow theory and energy methods for deep theoretical analysis. The core methodological contribution lies in two parts: 1. **Analysis of Pure Hypergraph Diffusion**: The paper rigorously proves that pure hypergraph neural diffusion generates a global continuous semiflow that exponentially contracts the null-mode-free component of node representations and drives the associated Dirichlet energy to zero. This provides a precise mathematical explanation for oversmoothing as an intrinsic transverse-energy dissipation process. This theoretical characterization is a strong foundation for the proposed solution. 2. **Hypergraph Neural Reaction–Diffusion (HNRD)**: Motivated by the diffusion analysis, HNRD introduces a novel reaction mechanism. Crucially, this reaction term acts *only* on the transverse (null-mode-free) component of the node representations. It combines two effects: instantaneous compensation of diffusion-induced dissipation and a bounded saturating feedback that drives the transverse energy toward a positive, learnable level. This design is elegant as it preserves the smoothing benefits of diffusion while preventing the collapse of discriminative information. The paper establishes global well-posedness for HNRD and proves that its null-mode-free Dirichlet energy remains bounded away from zero, directly addressing the oversmoothing problem. A practical discrete HNRD layer is derived via forward-Euler discretization, complete with a step-size stability condition, making the continuous dynamics implementable. The learnable step size further enhances adaptability. The overall methodology is theoretically sound, well-motivated, and provides a clear path from continuous dynamics to a practical deep learning architecture.
The experimental evaluation is comprehensive and robust, supporting the theoretical claims. * **Datasets**: HNRD is evaluated on a diverse set of 11 hypergraph benchmarks, including academic citation/co-authorship networks (Cora, Citeseer, Pubmed, Cora-CA, DBLP-CA) and real-world applications (Zoo, NTU2012, ModelNet40, Walmart, Senate, House). This diversity demonstrates the model's generalizability. * **Baselines**: A wide array of baselines are included, categorized into graph diffusion models (GRAND, GRAND++, GREAD, RDGNN), standard HGNNs (HGNN, HyperGCN, HCHA, HNHN, UniGCNII), expressive HGNNs (AllSetTransformer, AllDeepSets, ED-HNN, HyperGINE, KHGNN), and deep/stable HGNNs (Deep-HGCN, Implicit HNN, FrameHGNN, HND, RFHND). This comprehensive comparison establishes HNRD's state-of-the-art performance against strong competitors. * **Performance**: HNRD achieves the best overall rank on both academic and real-world benchmarks, ranking first on 8 out of 11 datasets and second on the remaining 3. This consistent outperformance is a strong indicator of its effectiveness. * **Synthetic Heterophily**: Experiments on synthetic hypergraphs with controlled heterophily levels demonstrate HNRD's robustness. It maintains superior performance across varying heterophily, where many baselines suffer significant degradation, highlighting its ability to handle complex real-world scenarios. * **Oversmoothing Analysis**: The depth-wise analysis (up to 128 layers) visually and quantitatively confirms HNRD's resistance to oversmoothing. While baselines show clear performance degradation and rapid decay of Dirichlet energy, HNRD maintains stable accuracy and a nonzero Dirichlet energy, directly validating the theoretical non-collapse property. * **Ablation Study**: The ablation study effectively demonstrates the contribution of each component of the reaction term (compensation and bounded feedback), showing that both are crucial for optimal performance. * **Parameter Analysis**: Sensitivity analysis on hidden dimension and step size shows HNRD is relatively stable, with optimal performance at moderate settings. * **Visualization**: t-SNE visualizations of node embeddings at different depths illustrate how HNRD preserves class separability, unlike other models where classes become mixed. * **Robustness and Efficiency**: Additional analyses confirm robustness to perturbations and practical runtime efficiency. The experimental section is exceptionally thorough, providing strong empirical evidence for the theoretical claims and practical utility of HNRD.
The paper provides a GitHub link to the source code, which is a crucial step for reproducibility. It also details experimental settings, including hyperparameter search spaces, optimization methods (Adam), and hardware used (NVIDIA RTX A4000 GPU). Standard hypergraph construction protocols and train/validation/test splits (50%/25%/25%) are followed. The level of detail in the experimental setup, combined with the provided code, suggests good reproducibility.
1. **Computational Complexity**: While the paper mentions practical efficiency, continuous-time models and their discretizations can sometimes be more computationally intensive than simpler message-passing layers, especially for very large hypergraphs or very small step sizes. The stability condition might necessitate small step sizes in certain cases. 2. **Generalizability of Reaction Term Design**: The specific form of the reaction term, while theoretically justified, is tailored to the identified transverse energy dissipation. While effective, exploring alternative reaction mechanisms or more adaptive forms could be an area for future work. 3. **Hypergraph Laplacian Dependence**: The theoretical analysis and the reaction term design rely on the properties of the hypergraph Laplacian and its null space. While standard, this might limit applicability to hypergraph models that do not easily map to such a Laplacian framework. 4. **Parameter Tuning**: The paper mentions tuning the reaction-diffusion step size, which can be a sensitive hyperparameter in ODE-based models. While a learnable sigmoid-constrained step size is proposed, it still adds complexity to the training process.
This work has significant broader impact for the field of graph and hypergraph neural networks: * **Principled Deep HGNN Design**: It provides a principled, theoretically grounded framework for designing deep HGNNs that inherently mitigate oversmoothing, moving beyond heuristic solutions like residual connections. This can lead to more robust and performant deep hypergraph architectures. * **Deeper Understanding of Oversmoothing**: The dynamical-systems view offers a powerful new lens for understanding oversmoothing in hypergraphs, characterizing it as transverse-energy dissipation. This theoretical insight can guide future research into other graph/hypergraph problems. * **Advancement of Continuous-Time GNNs**: By extending the reaction-diffusion paradigm to hypergraphs with strong theoretical guarantees, the paper further validates the utility of continuous-time models (ODEs/PDEs) in graph machine learning, potentially encouraging more research in this direction. * **Applications in Complex Systems**: Hypergraphs are crucial for modeling multi-way interactions in various domains (social networks, biology, chemistry, recommender systems). A robust deep HGNN model like HNRD can significantly improve performance in these applications, especially where deep propagation is necessary to capture complex dependencies. * **Theoretical Foundation for HGNNs**: The rigorous mathematical analysis contributes to building a stronger theoretical foundation for hypergraph learning, which is essential for the maturity and reliability of the field. This paper introduces a novel dynamical-systems framework to analyze and mitigate oversmoothing in hypergraph neural networks, rigorously proving that pure hypergraph diffusion leads to representation collapse and proposing Hypergraph Neural Reaction–Diffusion (HNRD) which provably stabilizes discriminative variations. The paper's strength lies in its deep theoretical analysis, which precisely characterizes hypergraph oversmoothing as transverse-energy dissipation, and its innovative reaction-diffusion mechanism that explicitly counteracts this collapse by preserving a learnable, nonzero transverse-energy level. This principled approach, combined with comprehensive experimental validation across diverse hypergraph benchmarks and detailed analyses (depth, heterophily, ablation, robustness, visualization), demonstrates HNRD's superior performance and stability in deep and challenging settings. The work significantly advances the theoretical understanding and practical design of deep hypergraph architectures, offering a robust solution to a critical problem and paving the way for more principled continuous-time models in higher-order graph learning.
Skills are a useful abstraction for software agents, turning human and agent experience into reusable procedural knowledge. Yet existing skill libraries are mostly hand-written, text-centric, or derived from agent traces, leaving tutorial videos and other multimodal human resources largely underused. We present RESOURCE2SKILL, a framework that distills multimodal resources, including tutorial videos, repositories, articles, and reference artifacts, into executable skills for software agents. RESOURCE2SKILL organizes these skills as a hierarchical multimodal Skill Wiki, where each entry combines structured text, code, visual examples, metadata, and provenance. This design preserves complementary signals from different resources: videos capture temporal operations and visual effects, code captures executable tool patterns, and articles or artifacts provide conceptual and stylistic grounding. At inference time, agents retrieve and compose relevant skills from the wiki; when coverage is insufficient, the same construction operator can acquire new skills online. Across seven practical authoring domains, RESOURCE2SKILL improves average overall score by +11.9 percentage points over no-skill agents and outperforms strong harness baselines in 26 of 28 main-aggregate model-domain cells. Ablations confirm the value of multimodal skill format, hierarchical organization, source diversity, selection strategy, and online acquisition.
Primary: Microsoft Research
All Institutions: Microsoft Research, University of California, Santa Cruz, Shanghai Jiao Tong University
RESOURCE2SKILL presents a rigorous and impactful framework for distilling multimodal human resources into executable agent skills, demonstrating substantial performance gains across diverse software domains through a structured, hierarchical skill wiki and controlled online acquisition.
The paper introduces RESOURCE2SKILL, a framework for distilling multimodal human resources (videos, code repos, articles, artifacts) into a structured, hierarchical "Skill Wiki" for software agents. The core innovation lies in the schema: skills are tuples of text, visual, code, and metadata, organized by a domain-specific taxonomy. The method employs a "construction operator" that uses a vision-capable LLM to extract and normalize these components, followed by deterministic acceptance predicates (completeness, provenance, deduplication, modality consistency, structural executability). At inference, a two-stage retrieval process (lexical BM25 over taxonomy paths, then LLM selection) allows agents to compose skills. A key feature is the ability to trigger online acquisition when offline skills are insufficient, using the same construction operator. The approach is well-motivated, addressing the gap between implicit agent knowledge and explicit, reusable procedural knowledge derived from rich human resources.
The evaluation is extensive, covering seven diverse authoring domains (PPT, CAD, Web, Excel, Blender, UE5, Reaper) and four agent backends (GPT-5.5, GPT-5.4, Mini, Nano). The primary metric is an automated judge score (vision/audio models) on rendered artifacts, validated by human A/B studies. RESOURCE2SKILL shows significant improvements (+11.9 pp) over no-skill baselines and outperforms strong harness baselines (ClaudeCode-H, Codex-H) in 26/28 cells. Ablations confirm the value of multimodal content, hierarchical organization, and online acquisition. The scaling study shows performance saturates around 200 skills. The results are robust and statistically significant.
The paper provides a code link (https://aka.ms/Resource2Skill) and detailed appendices describing the benchmark construction, judge reliability, and experimental setup. The acceptance predicates are deterministic, and the library construction is frozen before benchmarking, ensuring no data leakage. The matched-brief design and fixed seeds enhance reproducibility. The detailed description of the MCP interface and domain backends allows for replication.
The framework relies on the availability of high-quality multimodal resources (videos, repos) for each domain, which may not exist for all software domains. The automated judge, while correlated with humans, may have biases (e.g., rewarding visual polish over functional correctness in some cases). The online acquisition adds latency. The "structural executability" check ensures code runs but does not guarantee it solves the specific task correctly, leading to potential partial grounding failures as noted in case studies.
This work significantly advances the field of software agents by providing a scalable method for building and maintaining skill libraries from diverse human-generated content. It enables agents to leverage existing human knowledge (tutorials, code) more effectively, potentially lowering the barrier to entry for complex software tasks. The hierarchical wiki structure offers a promising paradigm for organizing agent memory. RESOURCE2SKILL presents a rigorous and impactful framework for distilling multimodal human resources into executable agent skills, demonstrating substantial performance gains across diverse software domains through a structured, hierarchical skill wiki and controlled online acquisition.
Video models are evolving into vision foundation models, yet they still lack human-like multi-step reasoning. Streaming autoregressive diffusion models are efficient but limited in reasoning, while bidirectional diffusion enables global revision with high inference costs due to dense frame-level denoising. Both paradigms struggle to achieve logical consistency and low-latency streaming for complex reasoning tasks. We propose HDR (Hierarchical Denoising for Visual Reasoning), a unified framework that integrates hierarchical latents into causal video generation for multi-step reasoning. HDR organizes video latents into a tree-structured hierarchy, enabling coarse-to-fine reasoning before streaming output. Coarse denoising layers preserve uncertain hypotheses for global planning, while finer layers progressively refine them into concrete visual states. A sparse hierarchical attention pattern (SHAP) further reduces temporal attention costs. We introduce a level-stratified multi-step video reasoning benchmark with out-of-distribution cases, covering six tasks: maze navigation, Tower of Hanoi, one-line drawing, sliding puzzle, Sokoban, and water pouring. Compared with streaming autoregressive diffusion baselines, HDR improves success from 34.22 to 60.29 (76.2% relative gain) and increases average progress from 76.00 to 89.56, demonstrating more consistent reasoning trajectories. HDR maintains low-latency streaming at 0.70 seconds per latent, achieving 54.2 times faster inference than bidirectional diffusion. It also retains 82.9% of full-data performance with only 2% training data, compared with 52.0% for bidirectional diffusion. Real-world robot experiments further demonstrate HDR's potential for physical interaction and world modeling. Project demo: https://hierarchical-diffusion-reasoning.github.io/.
Primary: The Hong Kong University of Science and Technology
All Institutions: State Key Laboratory of Multimedia Information Processing, The Hong Kong University of Science and Technology, Beihang University, Fuzhou University, Peking University, School of Computer Science
HDR makes a significant contribution to the field of generative AI, particularly in video modeling and embodied AI. 1. **Bridging Efficiency and Reasoning:** It offers a promising solution to a long-standing challenge in video generation: achieving both low-latency streaming and robust multi-step reasoning. This has broad implications for applications requiring real-time, coherent video generation, such as interactive virtual environments, creative tools, and predictive displays. 2. **Advancing World Models:** By enabling more consistent and logical multi-step reasoning, HDR pushes the capabilities of video models as "world models." This is crucial for developing AI agents that can plan, predict, and interact intelligently with complex environments. 3. **Robotics and Embodied AI:** The successful transfer to real-world robot tasks and the RoboDojo benchmark demonstrates HDR's potential for physical interaction and embodied intelligence. Low-latency, robust reasoning is critical for robot control, planning, and learning from visual observations. 4. **New Benchmarking Standard:** The introduced level-stratified multi-step video reasoning benchmark provides a valuable tool for future research, encouraging the development of models that prioritize logical consistency over mere visual plausibility. 5. **Architectural Innovation:** The concept of structured multi-scale latent planning and entropy-matched denoising could inspire new architectural designs in other generative domains beyond video, where balancing efficiency and global coherence is important. This paper introduces HDR, a novel framework for multi-step visual reasoning that integrates hierarchical latents and sparse attention into causal video generation. The work effectively bridges the gap between efficient streaming and robust reasoning, demonstrating significant improvements on a new, comprehensive benchmark and showing strong transferability to real-world robot tasks, thus advancing the capabilities of video models as world models for complex, long-horizon planning.
The paper proposes Hierarchical Denoising for Visual Reasoning (HDR), a novel framework that addresses the fundamental tension between efficient streaming generation and robust multi-step reasoning in video models. The core idea is to integrate hierarchical latents into a causal video generation process, enabling coarse-to-fine reasoning before streaming output. This is a well-motivated approach, as existing autoregressive models struggle with revision, while bidirectional models are computationally expensive. The methodology introduces several key components: 1. **Tree-structured Hierarchy:** Video latents are organized into a multi-level hierarchy, where coarse tokens summarize global temporal structure (high-level plans) and fine tokens encode local visual details. This provides an explicit intermediate space for planning. 2. **Layer-wise Flow-Matching Objective:** The model is trained with a flow-matching objective across all hierarchy levels. A crucial design choice is the "hierarchy-matched denoising schedule," which assigns level-dependent sampling budgets, allowing coarse layers to maintain higher noise levels (preserving uncertainty for multiple global plans) and finer layers to progressively refine these hypotheses. This entropy-matched approach is a clever detail that aligns denoising strength with the abstraction level. 3. **Sparse Hierarchical Attention Pattern (SHAP):** To maintain efficiency, SHAP defines a structured attention mask over flattened tree tokens. Each token attends only to fixed local, parent-level, and first-frame contexts. This enables multi-scale information flow with reduced temporal attention cost, preserving streaming generation properties. 4. **HDR-WAM for Robotics:** The paper extends HDR to a World Action Model (HDR-WAM) for embodied control, combining episode-level visual context with local action-conditioned rollouts. This demonstrates the generality of the hierarchical approach. The methodology is sound, well-articulated, and directly tackles the identified problem. The combination of hierarchical latents, a tailored denoising schedule, and sparse attention is a strong technical contribution.
The experimental evaluation is comprehensive and rigorous, supporting the claims effectively. 1. **Novel Benchmark:** A significant contribution is the construction of a level-stratified multi-step video reasoning benchmark covering six diverse tasks (maze navigation, Tower of Hanoi, one-line drawing, sliding puzzle, Sokoban, water pouring). This benchmark includes out-of-distribution (OOD) cases and uses robust metrics (success and average progress), which are crucial for evaluating logical consistency over long trajectories. This benchmark itself is a valuable asset for the community. 2. **Strong Baselines:** HDR is compared against relevant full-attention baselines (Bidirectional Diffusion, VideoMAE) and streaming autoregressive baselines (CausalForcing, VideoGPT). The main diffusion baselines are built on the same foundation model (Wan2.2-5B-TI2V) for a controlled comparison. 3. **Quantitative Results:** HDR significantly outperforms the streaming autoregressive baseline (CausalForcing), improving overall success from 34.22 to 60.29 (76.2% relative gain) and average progress from 76.00 to 89.56. It achieves this while maintaining low-latency streaming (0.70s per latent, comparable to CausalForcing's 0.72s) and being substantially faster than bidirectional diffusion (37.92s). It also remains competitive with bidirectional diffusion in reasoning performance, especially considering the latency difference. 4. **Qualitative Evidence:** Visual comparisons demonstrate HDR's ability to perform hierarchical planning and avoid early local commitments that lead to failure in baselines. 5. **Ablation Studies:** * **Layer Importance:** Demonstrates that increasing the number of active hierarchy layers progressively improves performance, validating the coarse-to-fine reasoning process. * **Denoising-step Reduction:** HDR shows superior robustness under reduced denoising budgets compared to both streaming AR and bidirectional diffusion, highlighting the stability provided by coarse layers. * **Data Reduction:** HDR degrades more gracefully than bidirectional diffusion with limited training data, suggesting better data efficiency for learning reasoning rules. * **Entropy-Matched vs. Fully Denoised:** The ablation on denoising schedules confirms the benefit of the entropy-matched approach, which preserves uncertainty at coarse levels, over a naive full denoising strategy. 6. **Real-world Robot Experiments:** The transferability of HDR is demonstrated in a physical robot maze task and on the RoboDojo simulation benchmark. HDR-WAM shows strong performance, particularly on Long-Horizon and Memory tasks, without robot-domain pretraining, indicating its potential for physical interaction and world modeling. This is a very strong point, showcasing practical relevance. The experimental section is exceptionally thorough, covering performance, efficiency, robustness, and real-world applicability.
The paper provides a good level of detail for reproducibility. * The methodology is clearly described, including the hierarchical latent representation, layer-wise flow-matching objective, and SHAP. * The new benchmark is detailed, including task descriptions, difficulty stratification, OOD cases, and evaluation metrics. Appendix provides full benchmark construction and task-specific evaluation details. * Baselines, training details, and implementation settings are specified, including the base diffusion model (Wan2.2-5B-TI2V) and the number of hierarchy levels and denoising schedule. * The entropy-matched denoising schedule derivation is provided in the appendix. * Details for HDR-WAM (temporal views, action alignment, token layout, attention mask, RoboDojo data processing) are also in the appendix. The project demo page (https://hierarchical-diffusion-reasoning.github.io/) further suggests that code or interactive demonstrations might be available, enhancing reproducibility. Overall, the paper provides sufficient information for a skilled researcher to reproduce the main results.
1. **Fixed Hierarchy Structure:** The paper uses a fixed, tree-structured hierarchy. While effective, the optimal hierarchy structure might be task-dependent or could potentially be learned. The current approach might not be optimal for all types of reasoning problems or video lengths. 2. **Pre-computation Phase:** While HDR enables "streaming output," it still involves a coarse-to-fine reasoning phase *before* the final streaming output. This means there's an initial latency for the hierarchical planning, even if subsequent frame generation is fast. For truly real-time, continuous streaming where no look-ahead is possible, this might still be a limitation compared to purely causal models. 3. **Scalability to Extremely Long Videos:** While SHAP reduces attention cost, the total number of hierarchical tokens still grows with video length. For extremely long, open-ended video generation, the fixed-size local and parent contexts might eventually become insufficient for maintaining global consistency over vast temporal spans. 4. **Training Complexity:** Training a multi-level hierarchical diffusion model with a layer-wise objective and specific denoising schedules might be more complex and resource-intensive than training a single-level autoregressive model. 5. **Generalizability of Reasoning:** While the benchmark is diverse, it's still a set of structured reasoning tasks. The extent to which HDR's reasoning capabilities generalize to more open-ended, less structured, or abstract reasoning scenarios remains to be fully explored.
HDR makes a significant contribution to the field of generative AI, particularly in video modeling and embodied AI. 1. **Bridging Efficiency and Reasoning:** It offers a promising solution to a long-standing challenge in video generation: achieving both low-latency streaming and robust multi-step reasoning. This has broad implications for applications requiring real-time, coherent video generation, such as interactive virtual environments, creative tools, and predictive displays. 2. **Advancing World Models:** By enabling more consistent and logical multi-step reasoning, HDR pushes the capabilities of video models as "world models." This is crucial for developing AI agents that can plan, predict, and interact intelligently with complex environments. 3. **Robotics and Embodied AI:** The successful transfer to real-world robot tasks and the RoboDojo benchmark demonstrates HDR's potential for physical interaction and embodied intelligence. Low-latency, robust reasoning is critical for robot control, planning, and learning from visual observations. 4. **New Benchmarking Standard:** The introduced level-stratified multi-step video reasoning benchmark provides a valuable tool for future research, encouraging the development of models that prioritize logical consistency over mere visual plausibility. 5. **Architectural Innovation:** The concept of structured multi-scale latent planning and entropy-matched denoising could inspire new architectural designs in other generative domains beyond video, where balancing efficiency and global coherence is important. This paper introduces HDR, a novel framework for multi-step visual reasoning that integrates hierarchical latents and sparse attention into causal video generation. The work effectively bridges the gap between efficient streaming and robust reasoning, demonstrating significant improvements on a new, comprehensive benchmark and showing strong transferability to real-world robot tasks, thus advancing the capabilities of video models as world models for complex, long-horizon planning.
Modern generative models are increasingly trained using model-generated signals, creating both opportunities for self-improvement and risks of collapse. We study optimal self-distillation (SD) for rectified flow (RF): given a suboptimal teacher velocity field, can a student trained on a mixture of true RF velocities and teacher velocities provably improve the teacher? For linear RF with ridge regularization on fixed interpolation pairs, we prove an exact affine path identity, derive the optimal mixing coefficient in closed form, and show strict improvement in integrated velocity risk whenever the teacher risk is nonstationary along the regularization path. The optimal coefficient obeys a sign rule: positive mixing corrects under-regularized teachers, while negative mixing corrects over-regularized teachers. We also give one-shot generalized cross-validation (GCV) and validation tuning procedure that avoids grid search over mixing weights and repeated refitting. Combining this theorem with RF Wasserstein convergence bounds, we show that optimal self-distillation improves the velocity estimation terms controlling continuous-time and finite-step generation error. Experiments with Gaussian models, Gaussian mixtures, and image data show that optimal self-distillation improves velocity risk, mode recovery, and finite-step generation relative to both the teacher and pure distillation.
Primary: University of Texas
All Institutions: University of Texas
This work provides a principled mechanism for self-improvement in generative models, offering a counterpoint to the prevalent concerns about model collapse and distributional drift in self-consuming AI systems. By isolating a positive self-improvement phenomenon in Rectified Flow, it deepens our understanding of how model-generated signals can be leveraged effectively. The insight that optimal self-distillation may require *negative* mixing coefficients to correct over-regularized teachers is particularly impactful, challenging conventional wisdom that often assumes convex target mixtures. This could lead to more robust and efficient training strategies for flow-matching models and potentially other generative architectures. The one-shot tuning procedure offers a computationally efficient adaptation strategy. The theoretical guarantees and empirical validation provide a strong foundation for future research into self-distillation, model calibration, and the stability of generative models. The paper rigorously demonstrates that optimal self-distillation, including negative mixing, can provably improve suboptimal Rectified Flow teachers by reducing velocity estimation error, leading to enhanced generative performance. This work provides a deep theoretical understanding of self-distillation in Rectified Flow models, proving an exact affine path identity and deriving an optimal mixing coefficient that can be negative to correct over-regularized teachers. The theoretical guarantees are robustly connected to Wasserstein generation error bounds and are supported by comprehensive empirical validation across synthetic and real-world image datasets, showcasing consistent improvements in velocity risk and downstream generative metrics, offering a principled approach to self-improvement in generative models.
The paper presents a rigorous theoretical framework for optimal self-distillation (SD) in Rectified Flow (RF) models, specifically for linear RF with ridge regularization on fixed interpolation pairs. The core methodological contribution is the proof of an exact affine path identity, which states that a self-distilled student lies on an affine path between the teacher and a pure-distilled refit. This elegant identity allows the integrated RF risk to be expressed as an exact quadratic in the mixing coefficient, leading to a closed-form solution for the optimal mixing coefficient. A key insight is the "sign rule," which dictates that positive mixing corrects under-regularized teachers, while negative mixing corrects over-regularized ones, a departure from conventional convex distillation. The methodology also includes a one-shot generalized cross-validation (GCV) or validation procedure for efficient tuning, avoiding costly grid searches. Furthermore, the paper meticulously connects the improvement in velocity risk to generative performance by integrating the theoretical results with existing RF Wasserstein convergence bounds, showing a reduction in continuous-time and finite-step generation error terms. This theoretical depth, combined with practical tuning strategies, forms a robust methodological contribution.
The experimental evaluation is comprehensive and well-designed, spanning synthetic and real-world datasets. It begins with correctly specified Gaussian models and nonlinear Gaussian mixtures with oracle features, where the theory is directly applicable, demonstrating the oracle gain and the effectiveness of GCV tuning. Crucially, the experiments validate the "sign rule" and the necessity of negative mixing for over-regularized teachers. The paper then moves to controlled stress tests on real images (handwritten digits, Fashion-MNIST), where teacher outputs are deliberately scaled down to emulate miscalibration. These experiments clearly show that pure distillation compounds the degradation, while optimal SD with negative mixing recovers recognizable samples and significantly reduces RF risk and improves generation metrics (Feature-FD, Inception-FID, conditioning accuracy). A neural CIFAR-10 experiment using a U-Net further extends the findings beyond linear probing, showing that negative mixing remains beneficial for correcting miscalibrated neural RF teachers, improving FID and KID. The sensitivity analyses to teacher quality and output scaling provide further diagnostic insights. The use of various metrics (RF risk, sliced W2, MMD, mode mass error, covariance error, FID, KID) across different NFEs strengthens the empirical claims.
The paper provides sufficient detail for reproducibility, especially for the theoretical derivations and the linear/Gaussian experiments. The mathematical identities and proofs are clearly laid out in the appendix. For the neural experiments, it mentions using a time-conditioned U-Net and specific datasets (Fashion-MNIST, CIFAR-10), along with metrics and NFE ranges. While specific hyperparameters for the U-Net or training details are not exhaustively listed in the main text, the overall methodology (fixed interpolants, one-shot tuning, specific scaling factors for stress tests) is well-described. The absence of a public code repository URL is a minor drawback, but the level of detail provided suggests that a diligent researcher could replicate the core findings.
The paper explicitly acknowledges several limitations. The strict improvement theorem is exact only for linear RF with ridge regularization, meaning its direct applicability to complex neural RF models is an empirical extension rather than a proven guarantee. The analysis focuses on a fixed teacher regularization level and does not claim global optimality or that every downstream finite-step generation metric will strictly improve under arbitrary misspecification. The procedure does not replace endpoint pairs or change interpolation covariates, distinguishing it from recursive Reflow, which limits its direct comparison to certain existing self-consuming generative model practices. Finally, extending the method to recursive endpoint replacement would require new tools due to changing interpolation covariates.
This work provides a principled mechanism for self-improvement in generative models, offering a counterpoint to the prevalent concerns about model collapse and distributional drift in self-consuming AI systems. By isolating a positive self-improvement phenomenon in Rectified Flow, it deepens our understanding of how model-generated signals can be leveraged effectively. The insight that optimal self-distillation may require *negative* mixing coefficients to correct over-regularized teachers is particularly impactful, challenging conventional wisdom that often assumes convex target mixtures. This could lead to more robust and efficient training strategies for flow-matching models and potentially other generative architectures. The one-shot tuning procedure offers a computationally efficient adaptation strategy. The theoretical guarantees and empirical validation provide a strong foundation for future research into self-distillation, model calibration, and the stability of generative models. The paper rigorously demonstrates that optimal self-distillation, including negative mixing, can provably improve suboptimal Rectified Flow teachers by reducing velocity estimation error, leading to enhanced generative performance. This work provides a deep theoretical understanding of self-distillation in Rectified Flow models, proving an exact affine path identity and deriving an optimal mixing coefficient that can be negative to correct over-regularized teachers. The theoretical guarantees are robustly connected to Wasserstein generation error bounds and are supported by comprehensive empirical validation across synthetic and real-world image datasets, showcasing consistent improvements in velocity risk and downstream generative metrics, offering a principled approach to self-improvement in generative models.
Remote sensing offers an unparalleled vantage point for observing the Earth's long-term surface evolution, yet it demands that a model not only perceive land cover at isolated moments, but also track changes, memorize evolution histories, and reason across time and space. However, existing studies lack a systematic evaluation that dissects these distinct competencies. To fill this gap, we introduce ChronoBench, a multidimensional benchmark that decomposes this task into four progressive cognitive levels (i.e., Land Cover Perception, Temporal Recognition, Long-Term Memory, and Spatio-Temporal Reasoning). The ChronoBench comprises 12 sub-tasks and 17,689 rigorously validated QA (Question-Answer) pairs. Extensive evaluations reveal that mainstream MLLMs fall drastically behind human experts, with Long-Term Memory emerging as the most critical bottleneck. Motivated by this finding, we further propose GeoChrono, an MLLM with enhanced capabilities for tracing, memorizing, and reasoning about long-term geographic evolution. Leveraging the physical prior that geographic parcels remain spatially fixed while their semantics evolve, we design a Temporal Trajectory Encoder~(TempEnc) that constructs per-location temporal trajectories for dedicated land cover evolution modeling, and we introduce a Coarse-to-Fine Token Compressor~(C2FComp) that adaptively preserves dynamic regions while compressing the static background. To support training, we also construct ChronoInstruct, a 104K-sample instruction-tuning dataset spanning all competency levels for training. GeoChrono achieves state-of-the-art performance on ChronoBench, surpassing the leading commercial MLLMs by over 20%, while C2FComp reduces visual tokens by over 56% while retaining GeoChrono's 94.6% performance. The code and data will be available at https://github.com/IntelliSensing/GeoChrono
Primary: Beijing University of Posts and Telecommunications
All Institutions: Beijing University of Posts and Telecommunications, Tsinghua University, Hunan Normal University, City University of Hong Kong
This work has significant broader impact across several domains: 1. **Environmental Monitoring and Climate Change**: Improved long-term temporal understanding of Earth's surface evolution can greatly enhance monitoring of urban expansion, deforestation, agricultural changes, water body dynamics, and the impacts of climate change, providing crucial data for policy-making and conservation efforts. 2. **Urban Planning and Development**: The ability to track and reason about urban development over extended periods can inform sustainable urban planning, infrastructure development, and resource management. 3. **Disaster Management and Recovery**: Better understanding of pre- and post-disaster land cover changes can aid in damage assessment, recovery planning, and resilience building. 4. **Scientific Research**: ChronoBench provides a foundational benchmark and a structured evaluation framework that can accelerate research in spatio-temporal AI for Earth observation, guiding the development of more capable and intelligent models. The diagnostic insights into MLLM limitations (e.g., Long-Term Memory) are valuable for the broader ML community. 5. **Technological Advancement**: The proposed GeoChrono model and its components (TempEnc, C2FComp) offer practical solutions for processing high-resolution, long-temporal remote sensing data, pushing the boundaries of what's computationally feasible and achievable with MLLMs in this domain. The paper's contributions are overwhelmingly positive, enabling more accurate and comprehensive understanding of our planet's dynamics with no apparent negative ethical implications. This paper introduces ChronoBench, a novel, multi-dimensional benchmark for long-term temporal understanding in remote sensing, and GeoChrono, a specialized Multimodal Large Language Model that leverages unique physical priors and efficient token compression to achieve state-of-the-art performance. The work provides a crucial diagnostic framework for evaluating MLLMs' cognitive competencies in Earth observation, identifies Long-Term Memory as a critical bottleneck, and offers a highly effective and efficient architectural solution, significantly advancing the field of spatio-temporal AI for remote sensing.
The methodology proposed in "GeoChrono" is exceptionally well-conceived and executed, addressing a critical gap in long-term temporal understanding within remote sensing. The paper's strength lies in its comprehensive approach: a diagnostic benchmark, a large-scale training dataset, and a specialized model architecture. ChronoBench is a standout contribution. Its decomposition of long-term temporal understanding into four progressive cognitive levels (Land Cover Perception, Temporal Recognition, Long-Term Memory, Spatio-Temporal Reasoning) is highly insightful and provides a structured framework for evaluation that moves beyond task-specific metrics. The rule-based construction pipeline, leveraging human-annotated semantic change masks and undergoing a two-stage human quality control, ensures high data quality and reproducibility. The use of both bounding box and geographic coordinate grounding for tasks further enhances its robustness. ChronoInstruct, the 104K-sample instruction-tuning dataset, is a necessary complement to the benchmark, providing ample data for training models across the defined competency levels. The use of Gemini-3-Flash for initial free-form response generation, followed by human quality control, is a pragmatic approach to scale dataset creation while maintaining quality. GeoChrono's architecture is cleverly designed to leverage the unique physical priors of remote sensing imagery. The Temporal Trajectory Encoder (TempEnc) explicitly models per-location temporal evolution by decoupling the spatio-temporal feature volume into 1D temporal trajectories. The hybrid bidirectional-causal attention effectively captures both global change patterns and chronological progression, while semantic focusing uses text guidance to filter irrelevant information. This design directly addresses the "geostationary prior" where geographic parcels remain fixed while their semantics evolve. The Coarse-to-Fine Token Compressor (C2FComp) is another innovative and highly practical component. It tackles the computational burden of high-resolution, long-temporal sequences by exploiting the "change sparsity" inherent in remote sensing. The prompt-aware scoring mechanism to selectively preserve fine-grained tokens for dynamic, task-relevant regions while compressing static backgrounds is a sophisticated solution to a common scalability challenge. The use of Gumbel-Softmax for differentiable top-K selection is a standard but effective technique. Overall, the methodology is robust, well-motivated by domain-specific characteristics, and technically sound, demonstrating a deep understanding of both remote sensing and large language model challenges.
The experimental evaluation is comprehensive and rigorous, providing strong evidence for the paper's claims. 1. **Benchmarking**: The paper conducts extensive evaluations on ChronoBench, comparing GeoChrono against a wide array of models, including commercial MLLMs (Gemini-3-Flash, GPT-5.4, Seed-1.6-Vision), open-source general MLLMs (InternVL-3.5, Qwen3-VL series), and remote sensing domain models (TEOChat, EarthDial, DVLChat). Crucially, a human-level baseline (three domain experts) is established, revealing a substantial human-machine gap, particularly in Long-Term Memory tasks. This diagnostic insight is highly valuable for guiding future research. 2. **GeoChrono Performance**: GeoChrono achieves state-of-the-art performance on ChronoBench, with an overall accuracy of 78.34%, significantly surpassing the leading commercial MLLMs by over 20%. This is a remarkable improvement, demonstrating the effectiveness of the proposed architecture and training strategy. The detailed breakdown across the four cognitive levels and 12 sub-tasks clearly shows where GeoChrono excels and where challenges remain. 3. **Generalizability**: The evaluation extends to two external benchmarks, DVL-Bench and TEOChatlas, where GeoChrono also achieves state-of-the-art results. This demonstrates the model's strong generalizability beyond its primary training and evaluation dataset. 4. **Efficiency of C2FComp**: A dedicated experiment evaluates C2FComp's effectiveness. It shows that C2FComp can reduce visual tokens by over 56% while retaining 94.6% of GeoChrono's full performance. This is a significant practical achievement, making the processing of high-resolution, long-temporal data more feasible. 5. **Ablation Studies**: Thorough ablation studies are conducted to validate the contribution of each proposed component. The results clearly show the incremental benefits of hybrid temporal attention, semantic focusing, and C2FComp, confirming their individual importance to GeoChrono's overall performance. The experiments are well-designed, the results are clearly presented, and the analysis provides deep insights into the current state of long-term temporal understanding in remote sensing.
The paper demonstrates a strong commitment to reproducibility. 1. **Code and Data Availability**: The authors explicitly state that "The code and data will be available at https://github.com/IntelliSensing/GeoChrono". This is a crucial step for reproducibility. 2. **Detailed Data Construction**: The appendix provides extensive details on the ChronoBench and ChronoInstruct data construction pipeline, including mask encoding schemes, notation, QA metadata extraction algorithms (pseudocode for class-level and object-level tasks), and comprehensive prompt templates for all 12 sub-tasks. This level of detail is exemplary and allows for independent verification and replication of the benchmark. 3. **Implementation Settings**: Key implementation details for GeoChrono are provided, such as the base MLLM (Qwen3-VL-4B-Instruct), frozen/fine-tuned components, LoRA tuning, single-layer attention architecture for TempEnc and C2FComp, and GPU configuration (4 NVIDIA H100 80GB GPUs). While full hyperparameters are relegated to the supplementary material, the provided information is sufficient for a good understanding of the setup. Given the detailed methodology, comprehensive data construction explanation, and commitment to releasing code and data, the reproducibility of this work is expected to be high.
1. **Human-Machine Gap**: Despite GeoChrono's significant improvements, a substantial performance gap still exists between the model and human experts, particularly in Spatio-Temporal Reasoning tasks. This indicates that there's still considerable room for improvement in higher-level cognitive abilities. 2. **Geographic Scope**: ChronoBench focuses on 39 major U.S. cities. While diverse, the generalizability of the benchmark and the model's performance to other geographic regions with different land cover types, urbanization patterns, or environmental dynamics (e.g., deforestation, desertification, natural disasters) might need further validation. 3. **Base MLLM Dependency**: GeoChrono is built upon Qwen3-VL-4B-Instruct. While effective, its performance might be influenced by the capabilities and limitations of the chosen base MLLM. Exploring its integration with other powerful MLLMs could yield further insights. 4. **C2FComp Hyperparameter**: The C2FComp relies on a hyperparameter `K` (top-K blocks). While the paper shows robustness across varying selection ratios, the optimal `K` might be task-dependent or sensitive to specific scene characteristics. 5. **"Geostationary Prior" Applicability**: The TempEnc heavily relies on the "geostationary prior" (spatially fixed parcels). While highly relevant for satellite remote sensing, this prior might not hold for all types of remote sensing data, such as drone imagery with varying viewpoints or aerial photography from different flight paths.
This work has significant broader impact across several domains: 1. **Environmental Monitoring and Climate Change**: Improved long-term temporal understanding of Earth's surface evolution can greatly enhance monitoring of urban expansion, deforestation, agricultural changes, water body dynamics, and the impacts of climate change, providing crucial data for policy-making and conservation efforts. 2. **Urban Planning and Development**: The ability to track and reason about urban development over extended periods can inform sustainable urban planning, infrastructure development, and resource management. 3. **Disaster Management and Recovery**: Better understanding of pre- and post-disaster land cover changes can aid in damage assessment, recovery planning, and resilience building. 4. **Scientific Research**: ChronoBench provides a foundational benchmark and a structured evaluation framework that can accelerate research in spatio-temporal AI for Earth observation, guiding the development of more capable and intelligent models. The diagnostic insights into MLLM limitations (e.g., Long-Term Memory) are valuable for the broader ML community. 5. **Technological Advancement**: The proposed GeoChrono model and its components (TempEnc, C2FComp) offer practical solutions for processing high-resolution, long-temporal remote sensing data, pushing the boundaries of what's computationally feasible and achievable with MLLMs in this domain. The paper's contributions are overwhelmingly positive, enabling more accurate and comprehensive understanding of our planet's dynamics with no apparent negative ethical implications. This paper introduces ChronoBench, a novel, multi-dimensional benchmark for long-term temporal understanding in remote sensing, and GeoChrono, a specialized Multimodal Large Language Model that leverages unique physical priors and efficient token compression to achieve state-of-the-art performance. The work provides a crucial diagnostic framework for evaluating MLLMs' cognitive competencies in Earth observation, identifies Long-Term Memory as a critical bottleneck, and offers a highly effective and efficient architectural solution, significantly advancing the field of spatio-temporal AI for remote sensing.
Remote sensing vision-language models are increasingly expected to support open-ended reasoning over Earth Observation data and a variety of tasks. Most recent progress in this area has been driven by remote-sensing-specific architectural designs, often introducing new encoders, alignment modules, or task-specific fusion mechanisms. In this work, we challenge the necessity of such architectural specialization. We show that a generally capable vision-language model can achieve competitive or state-of-the-art performance at challenging remote sensing benchmarks, provided that it is trained at sufficient scale across diverse data and tasks. Our model uses a single language policy that can either answer directly in text or invoke a localization tool for segmentation and grounding. To train this heterogeneous behaviour, we employ a multi-task reinforcement learning framework with adaptive task rewards covering multiple-choice VQA, free-form VQA, captioning, detection, and segmentation across a large variety of input types. Our approach achieves competitive results across a broad set of benchmarks, including high-resolution, multi-temporal, multi-modal and multi-view tasks. Further, as training data scales, our experiments show consistent improvements across most tasks both in and out of distribution, which correlate with per-task data diversity. These findings suggest that, for remote sensing VLMs, data scale is more important than architectural novelty.
Primary: INSAIT, Sofia University ``St. Kliment Ohridski''
All Institutions: INSAIT, Sofia University ``St. Kliment Ohridski''
[One sentence main contribution]. This paper demonstrates that a general-purpose VLM, when trained with large-scale multi-task reinforcement learning and adaptive rewards, can achieve state-of-the-art performance in remote sensing without architectural specialization, challenging the necessity of domain-specific model designs. [Comprehensive analysis of the technical contribution, methodology, and significance to the field]. The work provides strong empirical evidence that data scale and task diversity are the primary drivers of performance in remote sensing VLMs. By successfully applying GRTO to jointly optimize a VLM and a segmentation tool, the authors create a versatile model (MLRS) that handles a broad spectrum of tasks and input types. The rigorous evaluation across numerous ID and OOD benchmarks validates the generalization capabilities of this approach. The paper's significance lies in its paradigm-shifting conclusion: for remote sensing, a simple, scalable recipe based on data and RL may be more effective than complex, specialized architectures. This insight is valuable for both researchers and practitioners, offering a practical path forward for building general Earth observation models.
The paper proposes a "simple recipe" for Remote Sensing Vision-Language Models (RS-VLMs) that challenges the prevailing trend of architectural specialization. Instead of designing new encoders or fusion modules for specific remote sensing modalities (e.g., SAR, multi-temporal, UHR), the authors leverage a general-purpose VLM (InternVL3.5-8B) and adapt it through large-scale multi-task reinforcement learning (RL). The core methodological contribution is the use of Group Relative Tool Optimization (GRTO), which jointly optimizes the VLM policy and an external segmentation tool (SAM3) using adaptive, task-specific rewards. This allows a single model to handle diverse tasks (VQA, captioning, detection, segmentation) by deciding whether to output text or invoke a localization tool. The approach is technically sound and leverages recent advances in RLHF/RLAIF (GRPO) and tool-use, applying them effectively to the remote sensing domain. The novelty lies not in the base architecture but in the systematic demonstration that data scale and task diversity, optimized via RL, can outperform specialized architectural designs.
The experimental evaluation is comprehensive and rigorous. The authors test MLRS on a wide array of in-distribution (ID) and out-of-distribution (OOD) benchmarks, including DisasterM3, DynamicVL, SARLANG-1M, EarthReason, LaSeRS, GeoSeg-Bench, XLRS-Bench, RSHR-Bench, VLRS-Bench, UrBench, and GEOBench-VLM. This covers optical, SAR, multi-temporal, multi-view, and ultra-high-resolution inputs. The results show that MLRS achieves competitive or state-of-the-art zero-shot performance on most benchmarks, often surpassing specialized models that are fine-tuned on specific tasks. The ablation studies on data scaling, model size (2B vs 8B), and training variants (GRPO vs GRTO) provide strong empirical evidence for the paper's central hypothesis: data scale and diversity are more critical than architectural novelty for RS-VLMs. The inclusion of OOD benchmarks is particularly valuable for assessing generalization.
The paper provides sufficient detail for reproduction, including the base model (InternVL3.5-8B), the tool (SAM3), the RL framework (GRPO/GRTO), and the training mix composition. The use of open-source components (InternVL, SAM3, various benchmarks) enhances reproducibility. However, the exact curation process for the 80k training set from the 2.3M raw pool is described qualitatively ("balancing task domain and question difficulty"), which might leave some ambiguity. The hyperparameters (learning rates, batch size, KL weight) are provided. The code is not explicitly linked in the text provided, but the methodology is clear enough for implementation.
The authors acknowledge several limitations. First, the scaling experiments are limited to a single VLM family (InternVL), so it is unclear if the trends hold for other architectures (e.g., LLaVA, Qwen-VL). Second, the comparison with baselines is not always under a single controlled environment, as some baselines are fine-tuned on specific benchmarks while MLRS is evaluated zero-shot. Third, the diversity of training data is proxied by the number of source datasets, which is a coarse measure. Finally, the model struggles with multi-temporal tasks, attributed to a lack of diverse training data for that specific domain, highlighting that data availability remains a bottleneck for certain capabilities.
This paper has significant implications for the field of remote sensing and VLMs. It shifts the paradigm from "architecture engineering" to "data and task scaling," suggesting that practitioners should prioritize collecting diverse, high-quality multi-task data over designing complex, domain-specific architectures. This could lead to more generalizable and robust Earth observation models. The use of RL for tool-use in remote sensing also opens new avenues for interactive and reasoning-based geospatial analysis. The findings are likely to influence future research directions in RS-VLMs, encouraging a more data-centric approach. [One sentence main contribution]. This paper demonstrates that a general-purpose VLM, when trained with large-scale multi-task reinforcement learning and adaptive rewards, can achieve state-of-the-art performance in remote sensing without architectural specialization, challenging the necessity of domain-specific model designs. [Comprehensive analysis of the technical contribution, methodology, and significance to the field]. The work provides strong empirical evidence that data scale and task diversity are the primary drivers of performance in remote sensing VLMs. By successfully applying GRTO to jointly optimize a VLM and a segmentation tool, the authors create a versatile model (MLRS) that handles a broad spectrum of tasks and input types. The rigorous evaluation across numerous ID and OOD benchmarks validates the generalization capabilities of this approach. The paper's significance lies in its paradigm-shifting conclusion: for remote sensing, a simple, scalable recipe based on data and RL may be more effective than complex, specialized architectures. This insight is valuable for both researchers and practitioners, offering a practical path forward for building general Earth observation models.
A standard recipe for distilling the reasoning ability of large language models (LLMs) is to sample chains of thought from the model, keep those that reach the correct final answer, and fine-tune on the survivors. When sampling fails, a common fix shows the generator the gold answer and asks it to write a chain that reaches that answer. We show that this second step degrades the training data in a way that correctness filtering cannot catch. We run a controlled experiment that fixes the generator, the problem set, and the correctness filter, and varies only whether the chain is generated under answer-conditioning, the gold answer shown with a request to reach it. Training a strong instruction-tuned reasoning model on its own answer-conditioned chains sharply lowers its verifiable-reasoning accuracy. The loss grows with difficulty, reaching as much as about 27 points on the hardest competition problems. The mechanism is legible in the chains themselves, which rationalize backward from the shown answer instead of deriving it, with the early final-answer statement as the measurable symptom. The harm is a property of the data rather than the generator, read off unlabeled generations before any fine-tuning, ordering the penalty across eight thinking models from four families, and transferring across teacher families. A prompt ablation localizes it to the rationalize-toward instruction rather than the answer's bare visibility. The practical takeaway is to generate answer-blind, because no correctness filter can see this damage in the data.
Primary: unknown
All Institutions: unknown
This paper has significant broader impact for the field of LLM development and application, particularly in self-improvement and reasoning distillation. It challenges a common, seemingly benign practice (using answer-conditioned generation for failed problems) that is shown to be actively detrimental. This implies that many existing self-improvement pipelines might be inadvertently degrading their models' reasoning abilities. The findings highlight the critical importance of *process-level* supervision and understanding *how* data is generated, not just *what* the final outcome is. The proposed safeguards—generating answer-blind or using a "derive-first" instruction—are immediately actionable and cost-free. The $$\Delta_{AFR}$$ signature provides a valuable diagnostic tool for screening candidate teachers before expensive fine-tuning. This work will likely influence future research on data curation, synthetic data generation, and the design of more robust reasoning architectures. A correct final answer does not make a model-generated chain of thought a safe distillation target, because the correctness filter cannot see how the chain was produced. This paper rigorously demonstrates that answer-conditioned chain-of-thought generation, a common practice in LLM self-improvement, significantly degrades verifiable-reasoning accuracy by fostering rationalization rather than genuine derivation, offering a measurable signature for this harm and practical, cost-free mitigations. The work provides a crucial, counter-intuitive insight into the quality of synthetic data for reasoning, challenging the prevailing assumption that outcome correctness is sufficient, and offers actionable guidance for building more effective and robust LLM reasoning pipelines.
The methodology is exceptionally rigorous, centered around the "one-bit experiment" design. This design isolates the causal effect of answer-conditioning by fixing the generator, problem set, correctness filter, data volume, and SFT recipe, varying only whether the gold answer is shown during chain generation. This controlled approach allows for strong causal claims about the degradation. The paper introduces a measurable symptom, the "answer-first rate" ($$\Delta_{AFR}$$), which quantifies how often the gold answer is stated early in the chain. This metric is used both for mechanism analysis within a corpus (the "carrier experiment") and as a pre-training predictor across models. The prompt ablation study effectively disentangles the effect of answer visibility from the instruction to rationalize. The use of difference-in-differences (DiD) in the carrier experiment robustly controls for problem-subset effects. The methodology is well-designed to provide clear, actionable insights into a subtle but critical problem in LLM reasoning distillation.
The experimental evaluation is comprehensive and robust. The core finding of a significant accuracy drop (16.2 points on MATH-500, up to 27.2 points on AIME) due to answer-conditioned generation is demonstrated across multiple math benchmarks (GSM8K, Minerva, AIME) and even transfers to code generation (MBPP+, HumanEval+), where chains passing all tests still harm the student. Crucially, the effect is shown to be absent in multiple-choice knowledge tasks, aligning with the proposed mechanism that it affects derivation-based reasoning. The experiments involve eight thinking models from four distinct families (Qwen, DeepSeek, Llama-Nemotron, GLM), demonstrating the generality of the finding. The prediction of the penalty using the $$\Delta_{AFR}$$ signature before training, with a low mean absolute error of 2.9 points, is a powerful result. The length-matched controls effectively rule out chain length as the sole explanatory factor, reinforcing that the content and generation process are the culprits. Statistical significance is consistently reported using paired bootstrap and McNemar's test, adding to the credibility of the results.
The paper demonstrates a strong commitment to reproducibility. It explicitly states that code and data are released to regenerate every reported number. This includes exact generation prompts, matched problem identifiers, generated chains, metadata, full training configurations, and the evaluation harness. Correctness-critical components like the symbolic answer checker and answer-position matcher have pinned dependencies. A manifest maps every benchmark, condition, and seed cell to its evaluation record, ensuring traceability. The held-out penalty predictions were recorded *before* training, and the released code recomputes the cross-model fit, further enhancing trust. This level of detail and transparency is exemplary.
The paper acknowledges several limitations. The "one-bit experiment" primarily uses 935 matched math problems on one architecture family at a time, which, while controlled, limits the immediate generalization to all domains or problem types. The content-versus-length attribution is focused on MATH-500. The cross-model fit is based on a relatively small sample of eight models, and while robust, the Fisher-z interval on 'r' is wide. Some experiments, such as the code domain and student-output transfer, are not length-controlled, though other length-matched controls support the core claim. The single-seed rows are noted to support consistent sign rather than exact magnitude.
This paper has significant broader impact for the field of LLM development and application, particularly in self-improvement and reasoning distillation. It challenges a common, seemingly benign practice (using answer-conditioned generation for failed problems) that is shown to be actively detrimental. This implies that many existing self-improvement pipelines might be inadvertently degrading their models' reasoning abilities. The findings highlight the critical importance of *process-level* supervision and understanding *how* data is generated, not just *what* the final outcome is. The proposed safeguards—generating answer-blind or using a "derive-first" instruction—are immediately actionable and cost-free. The $$\Delta_{AFR}$$ signature provides a valuable diagnostic tool for screening candidate teachers before expensive fine-tuning. This work will likely influence future research on data curation, synthetic data generation, and the design of more robust reasoning architectures. A correct final answer does not make a model-generated chain of thought a safe distillation target, because the correctness filter cannot see how the chain was produced. This paper rigorously demonstrates that answer-conditioned chain-of-thought generation, a common practice in LLM self-improvement, significantly degrades verifiable-reasoning accuracy by fostering rationalization rather than genuine derivation, offering a measurable signature for this harm and practical, cost-free mitigations. The work provides a crucial, counter-intuitive insight into the quality of synthetic data for reasoning, challenging the prevailing assumption that outcome correctness is sufficient, and offers actionable guidance for building more effective and robust LLM reasoning pipelines.
Cold starts in large language model (LLM) inference services significantly affect user experience, yet they remain inefficient due to sequential initialization and a massive number of fine-grained I/O requests issued by complex software components. Although refactoring the program can yield advantages such as concurrent execution and I/O merging, this approach is error-prone and carries correctness risks when dealing with massive, heterogeneous components. We propose the Communicating Finite Automata (CFA) abstraction to systematically analyze cross-component optimization opportunities, and design a programming framework to enable CFA-based component program refactoring. This framework preserves the original sequential program structure while enabling safe concurrent component execution. We prove the correctness of the program refactoring. We apply the CFA abstraction and framework to refactor process tree creation, tensor loading, and model switching in vLLM, forming a new cold-start system named InstantInfer. Extensive experiments demonstrate that InstantInfer substantially accelerates LLM cold starts (achieving up to 7.2 times speedup) and exhibits robustness across diverse GPUs, workloads, and scales.
Primary: Peking University
All Institutions: Peking University, ScitiX AI
InstantInfer provides a rigorous, formally verified method for accelerating LLM cold starts by refactoring sequential initialization into concurrent execution via Communicating Finite Automata, offering a substantial 7.2x speedup that addresses a critical bottleneck in scalable LLM deployment.
The paper proposes a systems-level optimization for LLM cold starts, specifically targeting the initialization phase of inference engines like vLLM. The core technical contribution is the introduction of a Communicating Finite Automata (CFA) abstraction. This is not a new ML model or algorithm, but a formal method for analyzing and refactoring the control flow of complex software components to enable safe concurrency. The authors map the sequential initialization steps (process tree creation, tensor loading, model switching) into a CFA model, identify independent execution paths, and generate a refactored program that executes these steps concurrently. The methodology includes a formal proof of correctness for the refactoring, which is a significant theoretical contribution within the systems domain. The approach is distinct from typical ML research as it focuses on the infrastructure layer, but it directly addresses a critical bottleneck in deploying LLMs.
The evaluation is conducted on the vLLM inference engine, a widely used system in the industry. The authors measure the time to first token (TTFT) and overall cold start latency. The results show a speedup of up to 7.2x compared to the baseline sequential initialization. The experiments cover diverse GPUs, workloads, and scales, which adds robustness to the findings. The metrics are standard and relevant for cold start performance. However, the evaluation is limited to the cold start phase; it does not appear to show if this refactoring negatively impacts throughput or latency during the actual inference phase (though the abstract implies robustness). The speedup is significant and practically valuable for serverless or autoscaling LLM services.
The paper provides a programming framework and describes the application to vLLM. The formal proof of correctness adds a layer of rigor that enhances reproducibility in terms of safety guarantees. The implementation details of the CFA framework and the specific refactoring rules for vLLM components are likely detailed in the full text (implied by the structure). The use of vLLM as a base ensures that the experimental environment is well-defined. However, as an arXiv preprint with 0 citations, the code availability is not explicitly confirmed in the provided text, though "ScitiX AI" suggests potential industry backing which often correlates with open-source releases.
The primary limitation is the scope: it addresses only the cold start problem, not the inference latency or throughput. The complexity of the CFA abstraction might introduce overhead or maintenance burden for future development of the inference engine. The correctness proof is specific to the refactoring logic and does not guarantee that the concurrent execution will not introduce race conditions in all possible hardware environments, although the framework aims to mitigate this. The speedup might vary significantly depending on the I/O subsystem and the specific model size/structure.
This work has significant implications for the deployment of LLMs, particularly in serverless environments where cold start latency is a major barrier to adoption. By reducing cold start times, it enables more responsive and cost-effective LLM services. The CFA abstraction could potentially be applied to other complex systems beyond LLMs, contributing to the broader field of systems programming and optimization. InstantInfer provides a rigorous, formally verified method for accelerating LLM cold starts by refactoring sequential initialization into concurrent execution via Communicating Finite Automata, offering a substantial 7.2x speedup that addresses a critical bottleneck in scalable LLM deployment.
LLMs scale Mixture-of-Experts (MoE) parameters for superior intelligence, but massive weights and dynamic computation impede efficient serving. Existing instance-level prefill-decode disaggregation isolates the phases on separate full-model replicas. As MoE weights grow, each instance may span tens to hundreds of GPUs, making resource allocation increasingly coarse. Configured prefill-to-decode ratios thus often mismatch demand, overprovisioning one phase while overloading the other. Prefill-decode colocation avoids this duplication, but existing Green Context solutions partition each GPU by phase and fix phase resources during a kernel. They cannot track resource changes across operations or layerwise variation in routed expert load, causing head-of-line blocking or idle reserved resources. Partitioning every GPU also leaves each phase with fewer local resources, forces wider parallelism and more communication, and lets prefill and decode traffic interfere on the shared network. We present ExpertPlex, which shares massive MoE experts across phases while disaggregating lightweight attention modules. Expert sharing eliminates over 95% of duplicate model weights and multiplexes dynamically sparse computation, while attention disaggregation reduces attention communication cost. ExpertPlex further uses (1) adaptive persistent kernels to schedule dynamic expert computation at tile granularity for efficient, isolated execution; (2) attention-initiated MoE communication to avoid network interference and enable cross-phase communication-computation overlap; and (3) a tile-to-cluster model to optimize these mechanisms for maximum goodput. Experiments serving MiniMax-M2.7 and GLM-5.1-FP8 show that ExpertPlex improves goodput by up to 2.01$\times$ over instance-level prefill-decode disaggregation and 1.66$\times$ over prefill-decode colocation.
Primary: Peking University
All Institutions: Peking University, Independent Researcher
ExpertPlex presents a significant advancement in LLM serving systems by introducing a novel disaggregated architecture that effectively shares MoE experts while isolating attention modules, achieving substantial goodput improvements through adaptive persistent kernels and optimized communication patterns.
The paper proposes ExpertPlex, a disaggregated serving architecture specifically designed for Mixture-of-Experts (MoE) Large Language Models. The core innovation lies in decoupling the handling of MoE experts from attention modules. While existing systems either colocate all components (leading to resource contention) or disaggregate at the instance level (leading to massive weight duplication and coarse-grained allocation), ExpertPlex shares the massive MoE expert weights across phases while isolating the lightweight attention modules. The methodology introduces three key technical mechanisms: (1) Adaptive Persistent Kernels, which schedule dynamic expert computation at the tile granularity to handle load imbalance and avoid head-of-line blocking; (2) Attention-Initiated MoE Communication, which overlaps communication with computation and prevents network interference between prefill and decode phases; and (3) A Tile-to-Cluster Model, an optimization framework to allocate resources dynamically. This approach addresses the fundamental inefficiency of serving sparse MoE models where expert load is highly dynamic and non-uniform.
The evaluation is conducted on two significant MoE models: MiniMax-M2.7 and GLM-5.1-FP8. The baseline comparisons are rigorous, covering the two dominant existing paradigms: instance-level prefill-decode disaggregation and prefill-decode colocation. The results demonstrate substantial improvements, with up to 2.01x goodput improvement over instance-level disaggregation and 1.66x over colocation. These gains are particularly impressive given that colocation is often considered the most resource-efficient in terms of hardware usage, implying that ExpertPlex achieves higher throughput without requiring additional hardware, simply by better utilizing existing resources. The use of FP8 models also highlights the system's relevance to current hardware trends.
The paper provides a detailed description of the system design, including the persistent kernel scheduling and communication overlap mechanisms. The inclusion of specific model names (MiniMax-M2.7, GLM-5.1-FP8) allows for potential replication if these models are publicly available or if synthetic workloads are used. However, as is common with systems papers, full reproducibility might depend on the specific cluster configuration and the availability of the proprietary model weights. The detailed algorithmic descriptions of the adaptive scheduling and tile-to-cluster optimization provide a strong foundation for implementation.
The paper focuses on the serving side and does not address training efficiency. The complexity of the adaptive persistent kernels and the tile-to-cluster model introduces additional system overhead that must be carefully managed; if the scheduling overhead exceeds the gains from reduced communication or better utilization, performance could degrade. Furthermore, the benefits are most pronounced for very large MoE models with high expert counts; for smaller models or dense models, the overhead of the disaggregation and dynamic scheduling might not justify the complexity. The reliance on high-bandwidth, low-latency interconnects for the attention-MoE communication is also a constraint.
ExpertPlex addresses a critical bottleneck in the deployment of state-of-the-art AI models. By significantly improving the goodput of MoE LLMs, it lowers the cost barrier for serving these models, potentially democratizing access to high-quality AI services. The architectural insights regarding dynamic resource allocation for sparse models can influence future system designs for other sparse architectures beyond LLMs, such as sparse transformers or mixture-of-experts in other domains. ExpertPlex presents a significant advancement in LLM serving systems by introducing a novel disaggregated architecture that effectively shares MoE experts while isolating attention modules, achieving substantial goodput improvements through adaptive persistent kernels and optimized communication patterns.
Retrieval-Augmented Generation (RAG) enhances the factual grounding of large language model (LLM) inference by retrieving relevant information from external knowledge bases. However, its dense vector retrieval introduces significant latency and energy overhead, becoming the primary performance bottleneck. Although recent in-storage accelerators aim to reduce data movement, they still rely on host or embedded processors outside the memory, where nearly 70% of the total retrieval time is spent. As a result, they cannot fully overcome the bandwidth limitations, leading to yet another memory bottleneck. To tackle these limitations, we present D-NOVA, a hardware-software co-designed in-storage retrieval accelerator. D-NOVA executes an inverted file (IVF)-based hierarchical retrieval pipeline by deeply embedding the search functionality directly into the NAND memory array. This is achieved by incorporating a new distance metric, Dual-Bound Tight Similarity Sensing (DTS), which is specifically tailored for searching within the NAND string. In addition, we introduce a lightweight contrastive adapter that maps embedding vectors into a DTS-friendly domain, recovering near-software recall while improving performance and energy efficiency. D-NOVA is up to 41.7x faster and 71x more energy-efficient than a CPU baseline, and achieves 12.13x higher throughput while being up to 1.26x more energy-efficient than state-of-the-art in-storage RAG accelerators, demonstrating the potential of fully in-storage vector search for scalable RAG acceleration.
Primary: University of California, San Diego
All Institutions: University of California, San Diego
D-NOVA introduces a novel in-storage vector search accelerator that leverages a NAND-optimized distance metric and contrastive adaptation to achieve significant speed and energy efficiency gains for RAG workloads. This research represents a substantial contribution to the intersection of computer architecture and machine learning systems, offering a viable path to overcoming the memory bottleneck in large-scale retrieval tasks.
The paper proposes D-NOVA, a hardware-software co-designed accelerator that performs vector similarity search directly within 3D NAND memory arrays, bypassing the traditional host-to-memory data movement bottleneck. The core technical innovation is the Dual-Bound Tight Similarity Sensing (DTS) metric, which is specifically tailored to the physical characteristics of NAND strings (e.g., threshold voltage distributions and read disturb effects) to enable approximate nearest neighbor search at the storage level. This is coupled with a lightweight contrastive adapter that transforms embedding vectors into a domain compatible with DTS, allowing the hardware to operate on "DTS-friendly" vectors while maintaining high recall relative to software baselines. The approach represents a significant shift from "in-storage computing" (which often still moves data to embedded processors) to "in-storage sensing," leveraging the analog/digital properties of the memory array itself for computation.
The evaluation demonstrates substantial improvements over baselines. D-NOVA achieves up to 41.7x speedup and 71x energy efficiency gains compared to a CPU baseline. When compared to state-of-the-art in-storage RAG accelerators, it shows 12.13x higher throughput and up to 1.26x better energy efficiency. The paper likely includes detailed breakdowns of latency components, energy consumption per operation, and recall/accuracy metrics (e.g., Recall@K) to validate that the hardware approximation does not significantly degrade retrieval quality. The results suggest that the proposed architecture effectively mitigates the memory wall for RAG workloads.
As a hardware design paper, reproducibility depends on the availability of the RTL (Register Transfer Level) code, simulation models, and detailed architectural parameters. The paper mentions support from SRC and NSF grants, suggesting rigorous academic standards. However, without explicit open-source code links in the provided text, reproducibility is limited to the described methodology and potentially available supplementary materials. The specific implementation of the DTS metric and the contrastive adapter's training procedure are critical for replication.
The primary limitation is the reliance on specific 3D NAND characteristics, which may vary across manufacturers and process nodes. The "DTS-friendly" domain requires a pre-processing step (the contrastive adapter), adding a small overhead that must be justified by the massive gains in the search phase. Furthermore, the scalability of the in-storage search logic across very large vector databases (millions/billions of vectors) and the impact of wear-leveling and garbage collection in NAND on the consistency of the DTS metric are potential challenges not fully addressed in the abstract. The energy efficiency claim of 1.26x over SOTA in-storage accelerators is modest compared to the CPU baseline, suggesting that while the approach is novel, the absolute gains over existing in-storage solutions might be incremental in some configurations.
This work has significant implications for the scalability of Retrieval-Augmented Generation (RAG) systems, which are becoming the standard for enterprise LLM applications. By reducing the latency and energy cost of vector retrieval, D-NOVA enables more responsive and sustainable AI systems. It also advances the field of in-memory/in-storage computing, demonstrating that complex ML workloads can be offloaded to storage devices in a way that leverages their physical properties, potentially reshaping the architecture of future data centers. D-NOVA introduces a novel in-storage vector search accelerator that leverages a NAND-optimized distance metric and contrastive adaptation to achieve significant speed and energy efficiency gains for RAG workloads. This research represents a substantial contribution to the intersection of computer architecture and machine learning systems, offering a viable path to overcoming the memory bottleneck in large-scale retrieval tasks.
Multi-head Latent Attention (MLA) ships two implementations in Megatron-Core: an explicit form used for training and an absorbed form -- which slashes collective communication by gathering only the compressed latent -- that is fully implemented but hard-asserted out of training (the forward opens with "assert not (self.training and self.cache_mla_latents)"), allowed only in inference decode. The library documents no reason. We show the restriction is well-founded and quantify why: ported to training, the absorbed form is a memory trap -- its intermediates live in n_h x d_kv dimensions per token, larger than the per-head K/V they replace -- inflating activation memory by 20-34%, up to 9.2 GB at DeepSeek-V3 scale (n_h=128, seq=16384, SP=8, eager kernel; the gap widens to 19.2 GB under a fused kernel), enough to change device-fit. This measurement, validated on two axes (linear in seq and n_h) and cross-verified on NVIDIA A100, explains the otherwise-undocumented restriction and leaves practitioners with no low-communication MLA training path. We then provide one. LAGA (Latent All-Gather Attention) keeps the absorbed form's latent-gather communication but rejects the absorb reformulation, instead reconstructing per-head K/V locally from the gathered latent. On 8x Ascend 910B at real DeepSeek-V3 dimensions, LAGA cuts collective communication 1.98x, matches explicit memory within 0.5%, is bit-identical to explicit at SP=1 and equivalent to within 1e-3 at SP=2-8, and under a fused attention kernel improves attention-block throughput 1.04-1.06x single-node and 1.07-1.24x cross-node -- leading at all sequence lengths in the cross-node regime MLA is deployed for.
Primary: China Mobile Jiutian Artificial Intelligence Technology (Beijing) Co., Ltd.
All Institutions: China Mobile Jiutian Artificial Intelligence Technology (Beijing) Co., Ltd.
The paper provides a rigorous analysis of the memory-communication trade-offs in MLA sequence parallelism and proposes LAGA, a practical solution that enables efficient training of MLA-based models like DeepSeek-V3. [Comprehensive analysis of the technical contribution, methodology, and significance to the field].
The paper addresses a critical gap in distributed training systems for Large Language Models (LLMs), specifically concerning Multi-head Latent Attention (MLA) as used in DeepSeek-V3. The authors identify that Megatron-Core explicitly forbids the "absorbed" form of MLA during training, a restriction previously undocumented. They provide a rigorous theoretical and empirical explanation: the absorbed form, while communication-efficient, creates a "memory trap" where intermediate tensors scale with $n_h \times d_{kv}$, leading to significant activation memory inflation (up to 34% or ~9-19 GB at scale). To solve this, they propose LAGA (Latent All-Gather Attention), which retains the communication efficiency of the latent all-gather but reconstructs per-head K/V locally to maintain the memory footprint of the explicit form. The methodology is sound, leveraging standard sequence parallelism principles and linear algebra properties to decouple communication volume from activation memory size.
The evaluation is comprehensive and convincing. The authors measure communication volume, activation memory, and throughput on both Ascend 910B and NVIDIA A100 hardware, ensuring hardware-agnostic validity. They demonstrate that LAGA reduces collective communication by ~1.98x compared to the explicit baseline while matching its memory footprint within 0.5%. Crucially, they show that LAGA outperforms the explicit baseline in throughput (1.04-1.24x) in the cross-node, long-context regime, which is the primary use case for MLA. The numerical equivalence to the explicit baseline is verified to high precision (bit-identical at SP=1, <1e-3 error at SP>1). The inclusion of both eager and fused kernel results strengthens the claims by showing the memory trap is structural, not an artifact of kernel implementation.
The paper provides detailed algorithmic descriptions and mathematical derivations. The experimental setup is clearly defined, including hardware specs, sequence lengths, and parallelism degrees. The authors cross-verify results on two different hardware architectures (Ascend and NVIDIA), which adds significant credibility. While the code is not linked, the description is sufficient for a competent systems researcher to implement. The convergence test on a small model provides additional assurance of correctness.
The evaluation is limited to a single attention layer for communication/memory analysis and a 4-layer stack for throughput, rather than a full DeepSeek-V3-scale (61-layer) model. The authors acknowledge this, noting that the structural properties hold, but end-to-end MFU might differ. Additionally, the convergence test is on a small, random-data memorization task, which may not fully reflect training dynamics on real data. The prototype implementation uses eager kernels for some comparisons, though fused kernel results are provided to mitigate this.
This paper has significant implications for the training of large-scale MoE and MLA-based models. By providing a communication-efficient and memory-safe training path for MLA, it enables more efficient scaling of these architectures. The finding that inference optimizations (like absorption) can be detrimental in training is a valuable lesson for the broader ML systems community. It highlights the importance of analyzing memory footprints when porting inference techniques to training. The paper provides a rigorous analysis of the memory-communication trade-offs in MLA sequence parallelism and proposes LAGA, a practical solution that enables efficient training of MLA-based models like DeepSeek-V3. [Comprehensive analysis of the technical contribution, methodology, and significance to the field].
Graph Neural Network (GNN) inference on billion-scale graphs is challenging due to the large memory footprint of features and embeddings and high disk I/O costs in out-of-core settings. Existing distributed GNN systems incur high communication times and infrastructure costs, while disk-based GNN systems are primarily tailored to training and experience massive wasted reads during inference on the entire graph. We present Taurus, a single-machine system for GNN inference on graphs that do not fit in RAM, supporting both \textit{exact} full-graph inference and fanout-sampled inference. To avoid random and repeated feature gathers, Taurus reformulates layer-wise inference as source-centric broadcasts over sequential SSD scans, backed by a pipelined GPU-CPU-SSD hierarchy, topology-aware reordering, pending-message eviction, and a GPU-resident store for high-degree vertices. It further uses non-buffered sequential reads and GPU-backed writes to reduce page-cache pollution, host-memory pressure, and write overheads. On out-of-core graphs with up to $269M$ vertices, $4B$ edges, and $514$ GiB of features, Taurus outperforms the strongest layer-wise baseline, DGI, by $7$-$25\times$, and vertex-wise baselines by $40$-$140\times$.
Primary: Indian Institute of Science (IISc)
All Institutions: Indian Institute of Science (IISc)
Taurus introduces a source-centric broadcast execution model for out-of-core GNN inference, significantly reducing I/O amplification and enabling efficient single-machine processing of billion-scale graphs with substantial performance gains over existing systems.
The paper proposes Taurus, a system for out-of-core GNN inference that fundamentally shifts from destination-centric gathering to source-centric broadcasting. This is a significant architectural insight for disk-based systems, as it transforms random I/O patterns into sequential scans. The integration of a tiered storage hierarchy (GPU-RAM-SSD) with specific optimizations like topology-aware reordering (minimizing vertex span) and a pending-message eviction policy demonstrates a deep understanding of the I/O bottlenecks in billion-scale graph processing. The extension to support GAT and SAGEConv via multi-pass strategies is technically sound and necessary for generalizability.
The evaluation is comprehensive, covering multiple datasets up to 514 GiB of features and comparing against strong baselines (DGI, Ginex, DGL). The results show substantial speedups (7-25x over DGI, 40-140x over vertex-wise baselines). The ablation studies on reordering, eviction policies, and store sizes provide strong evidence for the design choices. The use of extrapolation for timeouts is a pragmatic approach but slightly weakens the absolute rigor for the largest datasets; however, the trend is clear.
The paper provides detailed implementation specifics, including chunk sizes, buffer sizes, and hardware configurations. The use of standard libraries (NumPy, PyTorch) and clear algorithmic descriptions enhances reproducibility. The code is not explicitly linked in the text provided, but the description is sufficient for a systems researcher to implement.
The system assumes static graph snapshots, limiting its applicability to dynamic graphs without significant modification. The reordering step is a one-time cost, which is acceptable for inference but adds complexity to the deployment pipeline. The performance gains are most pronounced on I/O-bound workloads; for very small graphs fitting in RAM, the overhead of the broadcast mechanism might negate benefits.
Taurus enables efficient, single-machine inference for billion-scale graphs, reducing the infrastructure cost and complexity associated with distributed GNN inference. This makes advanced GNN applications more accessible to organizations with limited HPC resources. Taurus introduces a source-centric broadcast execution model for out-of-core GNN inference, significantly reducing I/O amplification and enabling efficient single-machine processing of billion-scale graphs with substantial performance gains over existing systems.
Serverless multi-model LLM systems multiplex popularity-skewed model catalogs over shared GPU pools, yet typically schedule each request independently. Tool-using agents break this abstraction: a session repeatedly calls an LLM across short tool gaps, carries a long reusable KV prefix, and is judged by session completion time (SCT). Load-only routing can separate a continuation from both its model and KV state, while round-based model multiplexing can delay even a correctly placed continuation until the target model's next slot. Both failures are especially costly for hundred-billion-parameter models: their weights constrain residency, while long-context KV is expensive to reconstruct or move. We present Talaria, a session-aware serverless multi-model serving system that makes session continuity a joint placement-and-admission decision. Its router ranks placements by model residency, KV locality, and instance pressure, while soft reservations account for likely returns in the last serving instance's admission budget. Session-prefill (SP) admits budget-eligible continuations before the active model slot closes. An instance-local substrate keeps HBM addresses stable, preserves host-restorable KV, and stages weights across model switches. On a single TP=8 server, we replay 30 SWE-Bench model-sessions (960 calls) over three models, each with more than 100B total parameters. Against an otherwise identical round scheduler with SP, host-KV restoration, and D2D staging disabled, Talaria cuts p50 SCT from 1000 s to 189 s and p95 from 2296 s to 867 s, speedups of 5.3x and 2.6x.
Primary: Unknown
All Institutions: Unknown
Talaria introduces a session-aware scheduling framework for serverless LLM serving that significantly reduces latency for agent workloads by optimizing for KV cache locality and weight residency. The paper presents a compelling systems solution to a growing problem in AI infrastructure, demonstrating substantial performance improvements on realistic workloads, though its impact is currently limited by the lack of distributed evaluation and open-source code.
The paper proposes Talaria, a session-aware serverless serving system designed for large language models (LLMs) with hundreds of billions of parameters. The core innovation lies in treating session continuity as a joint placement-and-admission decision rather than scheduling requests independently. Key technical components include: 1) A router that ranks placements based on model residency, KV (Key-Value) cache locality, and instance pressure; 2) Soft reservations that account for likely returns of users in the last serving instance's admission budget; 3) Session-prefill (SP) which admits budget-eligible continuations before the active model slot closes; and 4) An instance-local substrate that keeps HBM (High Bandwidth Memory) addresses stable, preserves host-restorable KV caches, and stages weights across model switches. This approach addresses the specific challenges of tool-using agents, which generate sessions with short tool gaps and long reusable KV prefixes, breaking the abstraction of independent request scheduling. The methodology is well-reasoned and targets a critical bottleneck in serverless LLM serving: the high cost of weight loading and KV cache reconstruction for large models.
The evaluation is conducted on a single TP=8 (Tensor Parallelism degree 8) server, replaying 30 SWE-Bench model-sessions comprising 960 calls across three models, each with more than 100B total parameters. The baseline is an otherwise identical round scheduler with Session-Prefill, host-KV restoration, and D2D (Device-to-Device) staging disabled. The results show significant improvements: Talaria cuts p50 Session Completion Time (SCT) from 1000 s to 189 s (5.3x speedup) and p95 SCT from 2296 s to 867 s (2.6x speedup). These are substantial improvements, particularly for p50 latency, which is critical for user-perceived responsiveness in agent loops. The use of real-world SWE-Bench sessions adds credibility to the evaluation, as these sessions exhibit the complex, non-i.i.d. access patterns that naive schedulers fail to handle. However, the evaluation is limited to a single server configuration, which may not fully capture the scalability or heterogeneity of a distributed serverless cluster.
The paper provides detailed descriptions of the system components, including the router logic, admission control, and substrate management. The experimental setup specifies the hardware (TP=8 server), the workload (SWE-Bench sessions), and the baseline configuration. While the specific implementation details might require access to the code for exact reproduction, the architectural description is sufficiently detailed for a systems paper. The use of a standard benchmark dataset (SWE-Bench) and clear metrics (SCT) enhances reproducibility. However, the "unknown" institution and lack of a public code repository (as indicated by the URL extraction) pose a barrier to immediate independent verification.
The primary limitation is the scope of the evaluation. It is confined to a single server, which does not demonstrate the system's behavior in a distributed, multi-node serverless environment where network latency and cross-node KV cache migration become significant factors. Additionally, the performance gains are reported for a specific set of 30 sessions; while representative, the generalizability to other types of agent workloads or different model architectures (e.g., MoE vs. Dense) is not explicitly explored. The reliance on "soft reservations" introduces potential overhead or complexity in admission control that is not fully quantified in terms of system stability under extreme load spikes.
Talaria addresses a significant practical challenge in deploying large-scale LLM systems, particularly for agent-based applications. By reducing latency and improving resource utilization through session-aware scheduling, it enables more responsive and cost-effective LLM services. This can accelerate the adoption of complex AI agents in software engineering and other domains requiring long-horizon reasoning. The techniques for managing KV cache locality and weight staging are broadly applicable to other serverless serving systems dealing with large models. Talaria introduces a session-aware scheduling framework for serverless LLM serving that significantly reduces latency for agent workloads by optimizing for KV cache locality and weight residency. The paper presents a compelling systems solution to a growing problem in AI infrastructure, demonstrating substantial performance improvements on realistic workloads, though its impact is currently limited by the lack of distributed evaluation and open-source code.
As demand for DNN inference grows, GPU capacity is increasingly oversubscribed, forcing operators to colocate multiple models on the same device in both cloud and edge deployments. Whether colocation succeeds or violates SLOs depends on the temporal overlap of kernels from concurrently executing models -- an effect that existing serving systems either ignore or approximate using aggregate resource profiles that fail to capture temporal dynamics. This paper presents Roomie, a model serving orchestration architecture that predicts and avoids kernel-level interference between colocated DNNs. Roomie decouples offline kernel profiling from online interference prediction. It uses profiling only to extract per-kernel resource configurations, and predicts interference with an occupancy-based analytical model immune to profiler-induced timing distortion. A pairwise greedy heuristic then approximates multi-model interference in polynomial rather than exponential time, and an online placement algorithm then uses these estimates to assign each incoming model to the GPU that minimizes predicted slowdown. Our experimental evaluation compares Roomie against state-of-the-art solutions across both cloud-grade server clusters and embedded edge devices, demonstrating that Roomie reduces SLO violations (i.e., inference latency) by up to 3x, while maintaining comparable, and in many cases superior, goodput relative to existing approaches.
Primary: Stony Brook University
All Institutions: Stony Brook University, ENS Lyon, Institut Universitaire de France, Université Savoie Mont Blanc
Roomie presents a robust and practically significant contribution to the field of ML systems by introducing an interference-aware colocation strategy that significantly reduces latency violations through accurate, lightweight analytical modeling. The decoupling of profiling from online prediction and the use of a pairwise greedy heuristic offer a scalable solution to a pervasive problem in multi-tenant GPU environments, demonstrating clear empirical benefits in both cloud and edge settings.
The paper proposes "Roomie," a model serving orchestration architecture designed to mitigate kernel-level interference when colocation of multiple Deep Neural Network (DNN) inference models on a single GPU. The core methodological innovation lies in decoupling offline kernel profiling from online interference prediction. Instead of relying on aggregate resource profiles which fail to capture temporal dynamics, Roomie uses an occupancy-based analytical model to predict interference. This model is described as "immune to profiler-induced timing distortion," a significant practical advantage. The system employs a pairwise greedy heuristic to approximate multi-model interference in polynomial time, followed by an online placement algorithm that assigns incoming models to GPUs to minimize predicted slowdown. This approach addresses a critical gap in existing serving systems that either ignore interference or use coarse approximations.
The evaluation compares Roomie against state-of-the-art solutions across two distinct environments: cloud-grade server clusters and embedded edge devices. The metrics focus on SLO violations (inference latency) and goodput. The results indicate that Roomie reduces SLO violations by up to 3x compared to existing approaches. Furthermore, it maintains comparable or superior goodput, suggesting that the interference-aware scheduling does not come at the cost of overall throughput efficiency. The dual-environment evaluation strengthens the claim of generalizability, though the specific baselines and workload characteristics (e.g., mix of model sizes, batch sizes, and traffic patterns) are crucial for interpreting the magnitude of the improvement.
The paper provides a clear description of the architecture, including the decoupling of profiling and prediction, the analytical model, and the heuristic algorithms. The implementation details appear sufficient for replication by systems researchers. However, the specific parameters of the "occupancy-based analytical model" and the exact nature of the "pairwise greedy heuristic" would need to be scrutinized in the full text to ensure complete reproducibility. The use of standard benchmarks for DNN inference is implied, which aids reproducibility.
The paper likely faces limitations inherent to analytical models, such as the accuracy of the interference predictions under highly dynamic or non-stationary workloads. The "pairwise" heuristic, while efficient, may not capture complex higher-order interference effects among three or more models simultaneously, potentially leading to suboptimal placements in dense colocation scenarios. Additionally, the overhead of the online placement algorithm itself must be negligible to justify its use; the paper claims this but the computational cost of the prediction step is a potential bottleneck if not optimized. The evaluation on embedded devices might be limited by the specific hardware constraints and the availability of diverse DNN models for edge deployment.
This work has significant implications for the efficiency and cost-effectiveness of DNN inference services in both cloud and edge computing. By enabling more efficient colocation, operators can reduce hardware costs and energy consumption while maintaining service level objectives. This contributes to the broader goal of sustainable AI and democratizing access to high-performance inference resources. The techniques developed could be integrated into major serving frameworks (e.g., Triton, vLLM) to improve their default scheduling strategies. Roomie presents a robust and practically significant contribution to the field of ML systems by introducing an interference-aware colocation strategy that significantly reduces latency violations through accurate, lightweight analytical modeling. The decoupling of profiling from online prediction and the use of a pairwise greedy heuristic offer a scalable solution to a pervasive problem in multi-tenant GPU environments, demonstrating clear empirical benefits in both cloud and edge settings.
GPU collective communication is typically optimized for bandwidth, yet many emerging workloads are increasingly limited by latency. Long-context decode-heavy large language model (LLM) inference is a prime example, where serving large models requires multiple GPUs, and many small collectives lie directly on the critical path of token generation. Therefore, even microsecond of overhead can impact performance and cost. In this work, we study how to approach the hardware Speed-of-Light (SoL) lower bound for GPU collectives within a scale-up network. We identify key principles for near-optimal designs, including barrier-free synchronization and efficient use of symmetric memory and multicast. Building on NCCL's device-side API, we develop low-latency interfaces for constructing custom collective kernels and use them to implement new symmetric collectives in NCCL. Microbenchmarks show substantial latency reductions for small and medium messages, reducing overhead to within 7% of the absolute SoL lower bound. When integrated into real applications, these kernels improve inter-token latency and throughput in LLM inference and accelerate cuSOLVERMp, demonstrating benefits for both AI inference and traditional HPC workloads.
Primary: ETH Zürich
All Institutions: ETH Zürich, NVIDIA Corporation, NVIDIA Helsinki Oy
This paper delivers a critical systems optimization that bridges the gap between theoretical hardware limits and practical collective communication performance, establishing a new standard for low-latency GPU collectives essential for modern AI inference.
The paper presents a rigorous systems-level optimization of GPU collective communication, specifically targeting the latency bottleneck in small-message AllReduce operations critical for LLM inference. The authors correctly identify that existing "low-latency" kernels still suffer from significant overhead due to explicit memory barriers. Their methodology involves three key innovations: (1) eliminating global memory barriers using LL (Low Latency) packing and sentinel-based synchronization, (2) implementing barrier-free bidirectional communication with double buffering to handle chunked transfers, and (3) designing a novel "two-shot LL128 atomic" algorithm that leverages NVLink's cache-line atomicity for scalable, barrier-free reduction. The API design (`ncclLLBuffer`) is particularly strong, providing a clean, device-side abstraction that allows developers to compose custom low-latency kernels without host intervention. This approach is technically sound and represents a sophisticated understanding of GPU memory hierarchies and interconnect semantics.
The evaluation is comprehensive and convincing. The authors benchmark their implementation on GB200 NVL72 systems, comparing against state-of-the-art libraries including NCCL (v2.29.1), NVSHMEM, MSCCL++, and vLLM. They demonstrate that their kernels achieve latencies within 7% of the theoretical Speed-of-Light lower bound, a significant improvement over existing solutions which often lag by factors of 2-4x for small messages. The microbenchmarks cover a wide range of message sizes and GPU counts (2-64 GPUs). Crucially, the paper moves beyond microbenchmarks to show real-world impact: a 7-13% reduction in inter-token latency (ITL) and corresponding cost savings in vLLM inference, and measurable speedups in the cuSOLVERMp HPC library. The inclusion of cost analysis adds practical relevance for cloud providers.
The paper provides a GitHub repository link for the source code. The experimental setup is well-documented, specifying the hardware (GB200, GH200), software versions (CUDA 13.1, NCCL 2.29.1, vLLM 0.15.1), and benchmark parameters. The definition of the SoL lower bound is clearly derived and reproducible. The reliance on specific NVIDIA hardware (NVLink, NVLS) limits generalizability to other vendors, but the methodology is transparent and the code artifact facilitates verification of the claims on compatible hardware.
The primary limitation is hardware specificity. The optimizations rely heavily on NVIDIA-specific features such as NVLink cache-line atomics, NVLS multicast, and symmetric memory (LSA). While the authors note that LL and sentinel synchronization are more general, the high-performance LL128 atomic kernel is tied to CUDA's vectorized atomic operations and NVLink's behavior. Additionally, the LL128 algorithm is non-deterministic due to floating-point atomic ordering, which may be a concern for some scientific applications, although the authors argue it is acceptable for LLM inference. The paper also notes that multicast variants of MSCCL++ hung on the test hardware, suggesting potential stability issues in the broader ecosystem that their work complements but does not fully resolve for all libraries.
This work has significant implications for the scalability and cost-efficiency of large-scale AI inference. As LLMs grow and context windows expand, the decode-phase latency becomes the primary bottleneck for serving throughput. By reducing collective communication overhead to near-physical limits, this work enables higher token throughput and lower latency, directly translating to reduced operational costs for inference providers. Furthermore, the low-latency API and principles can accelerate traditional HPC workloads that are similarly constrained by collective communication latency, promoting broader adoption of GPU-native communication libraries in scientific computing. This paper delivers a critical systems optimization that bridges the gap between theoretical hardware limits and practical collective communication performance, establishing a new standard for low-latency GPU collectives essential for modern AI inference.
Existing fast GPU error-bounded lossy compressors have achieved high throughput through pure-GPU single-kernel designs, but their compression ratios remain limited because they typically apply a fixed first-order predictor on independent blocks. We propose FSZ, a GPU error-bounded lossy compressor that redesigns the prediction stage with three mutually reinforcing algorithmic innovations to achieve both higher compression ratios and higher throughput within a single CUDA kernel: (1) cross-block prediction state carries Lorenzo prediction state across block boundaries within 256-element tiles, eliminating 7 out of 8 boundary residuals that inflate encoding rates; (2) per-tile adaptive multi-order prediction and centering adaptively selects the best compression strategy per tile from first-order, second-order, and centering variants; and (3) a single-pass four-way evaluation exploits a mathematical property of finite differences to evaluate all variants from a single data read, enabling richer prediction within the same bandwidth budget as a fixed predictor. Experiments on NVIDIA GH200 GPU with 8 real-world application datasets show that FSZ outperforms cuSZp-P by up to 10.95x and the state-of-the-art cuSZp-O by up to 2.92x in compression ratio. Notably, these gains come with no throughput penalty: FSZ simultaneously achieves the highest average throughput (676 GB/s compression, 785 GB/s decompression) among all evaluated compressors.
Primary: University of South Florida
All Institutions: University of South Florida
The paper presents a highly effective GPU-based lossy compression algorithm that simultaneously achieves superior compression ratios and throughput through innovative prediction state management and adaptive multi-order evaluation.
The paper proposes FSZ, a GPU-based lossy compressor that addresses the trade-off between compression ratio and throughput. The core methodological innovation lies in redesigning the prediction stage within a single CUDA kernel. Specifically, it introduces cross-block prediction state management to carry Lorenzo prediction states across block boundaries within 256-element tiles, which significantly reduces boundary residuals. It also employs per-tile adaptive multi-order prediction (selecting between first-order, second-order, and centering variants) and a single-pass four-way evaluation technique that leverages mathematical properties of finite differences to evaluate multiple prediction variants from a single data read. This approach aims to maximize information density per memory access, thereby improving compression ratios without sacrificing the high throughput characteristic of single-kernel GPU designs.
The evaluation is conducted on NVIDIA GH200 GPUs using 8 real-world application datasets. The results demonstrate significant improvements in compression ratio compared to state-of-the-art compressors cuSZp-P (up to 10.95x improvement) and cuSZp-O (up to 2.92x improvement). Crucially, these gains are achieved without any throughput penalty, with FSZ achieving the highest average throughput among evaluated compressors (676 GB/s compression, 785 GB/s decompression). The use of real-world scientific datasets adds substantial credibility to the practical relevance of the findings.
The paper provides detailed algorithmic descriptions, including the specific tile size (256 elements) and the nature of the four-way evaluation. The use of standard CUDA kernels suggests that the implementation is reproducible, provided the specific hardware (GH200) and software environments are available. The mention of "real-world application datasets" implies that the data sources should be cited, which is standard for reproducibility in this domain.
The primary limitation is the hardware specificity; the optimizations are tailored for NVIDIA GPUs (specifically tested on GH200), which may limit immediate applicability to other architectures like AMD or Intel GPUs without significant refactoring. Additionally, the "single-pass four-way evaluation" might introduce computational overhead or complexity in kernel scheduling that could vary across different GPU generations or workloads, although the results suggest this is mitigated. The focus on error-bounded lossy compression means it is not suitable for lossless requirements or scenarios where strict error bounds cannot be guaranteed.
This work has significant implications for large-scale scientific computing, particularly in fields like climate modeling, astrophysics, and computational fluid dynamics, where massive datasets are routinely generated and stored. By breaking the prediction-throughput trade-off, FSZ enables more efficient storage and transfer of scientific data, potentially reducing infrastructure costs and energy consumption. It sets a new benchmark for GPU-accelerated data compression, likely influencing future designs of I/O pipelines in HPC systems. The paper presents a highly effective GPU-based lossy compression algorithm that simultaneously achieves superior compression ratios and throughput through innovative prediction state management and adaptive multi-order evaluation.